The Windfall Clause has a remedies problem

John Bridge 🔸

[NB: I shelved this project in 2023 due to competing research demands. If you would like to speak about anything I've written, feel free to reach out directly]

Thanks to Holly Scott, Aryan Yadav, Jide Alaga, Will Greenman, Cullen O’Keefe, Haydn Bellfield and Peter Wills for all your feedback and suggestions on this post.

This post is a supplement to the (forthcoming) TaWWWC - England & Wales, a discussion of the viability of the Windfall Clause in English law. That post is, itself, a part of ‘Towards a Worldwide, Watertight Windfall Clause’, a sequence I’m writing on the legal viability of the Windfall Clause in seven important common law jurisdictions.

Today, I’m writing about what I believe is a serious issue with the Windfall Clause - a lack of satisfactory remedies in the event of a Developer's breach of contract. By flagging this issue and raising possible solutions before the Clause is widely adopted, I hope to encourage critical engagement with the question of whether the Clause is, in fact, good longtermist policy. If our answer to that question remains affirmative, I hope to stimulate further discussion of how we can craft a truly watertight Windfall Clause.

How to read this post:

I’m aiming for this post to work as a standalone piece, so I have a few suggestions to make sure you understand everything I’ve written below:

Check the defined terms - I’ve used several defined terms throughout this sequence to improve readability. Make sure you have read the definitions in my original post beforehand or you will be confused.
Open a second window with my recommendations - I link back to my recommendations throughout this post. You'll save yourself a headache if you can quickly flick over to them as you read.
Feel free to skip ‘The Law’ - Unless you’re looking for a more thorough grounding in the law on damages and specific performance, you can probably skip straight to my discussion of the consequences for the viability of the Clause.
If you’re not a lawyer - Unless you’re feeling masochistic, feel free to skip most of this post and focus on the takeaways and recommendations. You may also want to check out my note on legalese and note on numbers to understand why some of the below is so inaccessible.
If you’re not familiar with the Clause - Watch this explainer by Cullen O’Keefe or read the original report to get a better understanding of how it’s supposed to work.

Finally, please note that I have only considered the availability of remedies in an English legal context here. The rules on remedies are broadly similar across all common law systems, so if there is an issue in English law it’s likely that similar issues exist in other key jurisdictions. However, there’s a chance that my criticisms don’t generalise, so I invite comments and feedback from those with more expertise in other legal systems to help determine how widespread this issue is.

Takeaways and recommendations:

For many plausible breaches of the terms of the Agreement, the Counterparty will not have access to a satisfactory remedy in English courts. This lack of remedies has two important implications for the viability of the Clause:

The availability of efficient breach - All things considered, it will likely be cheaper in expectation for a Developer of WGAI to breach the contract rather than bear the costs of complete performance. This means that a rational Developer will be very likely to default on the Agreement.
A lack of remedies - In the event that the Developer does decide to breach its obligations, the courts are unlikely to provide equitable remedies or damages proportionate to the scale of the Counterparty’s losses in expectation. This means that the Counterparty will not receive a satisfactory remedy for the Developer’s breach.

I do not see any foolproof methods to resolve these issues. However, the drafters of the Agreement can take steps to improve the enforceability of the Clause. My recommendations for this are as follows:

Refer any disputes about the likelihood of WGAI to expert determination - To avoid issues with poorly-calibrated judges, the drafters may wish to insert a provision which provides that any calculation of damages for the Developer’s breach would be subject to a binding valuation by experts who are predetermined by the parties.
Insert a basis clause ^[1] in the Agreement - The courts might uphold a provision stating that the Developer and Counterparty are contracting on the basis that the Developer achieving WGAI is not ‘remote speculation’. If so, this would prevent the Developer from asserting the opposite in court in order to argue that the Counterparty’s losses are too remote. In theory, this would increase the damages available to the Counterparty.
Do not draft a purely donative Agreement - This is essential to avoid any issues with a lack of equitable remedies due to failure of consideration.
Include a provision stating that specific performance is the primary remedy for breach of the Clause - This might help ward off arguments by the Developer that damages is an adequate remedy for breach of the Clause. Alternatively, it could function as a bluff clause to discourage a Developer’s initial breach.^[2]
Strongly consider a shares-based Clause - Courts are more likely to provide specific performance if the Developer has covenanted to issue a unique class of shares which the Counterparty couldn’t acquire elsewhere.
Seek a Counsel’s Opinion on the issue with remedies - I am a law student, not a lawyer, so my analysis may be missing key considerations. This issue seems serious enough to merit the cost of further investigation by an experienced commercial barrister.

Outstanding questions:

Even if these suggestions are adopted, there are several issues which still need to be addressed before the Windfall Clause is implemented:

Will English courts uphold a basis clause? There’s a significant risk that the courts will strike out a basis clause of the sort outlined below on the grounds that it is a de facto penalty clause. Given that a basis clause could prove central to the Counterparty receiving adequate damages, further research into the enforceability of such a clause seems important.
Will an expert determination clause provide proportionate damages? There’s a risk that pre-determined experts will become poorly-calibrated over time, or that the Developer will insist on including WGAI sceptics on any expert panel. The level of disagreement between experts also indicates that the size of any award could vary wildly depending on which experts are chosen. As such, significant consideration is needed to decide exactly who will constitute the panel of experts if an expert determination clause is to be used.
Is this a problem across common law jurisdictions? If this is only a problem in English law, then it should be possible to sidestep the issue by stipulating that the Agreement be governed by foreign law. It will also be important to determine which jurisdictions will enforce whose foreign laws, and how consistently they do so.^[3]
Under what conditions can the Developer engage in efficient breach? This is not a legal question per se, but it is important for determining the viability of the Clause. Further research is needed to determine the expected value of default for a Developer, and how the likelihood of enforceability of the Clause affects this.
Is a court more likely to order specific performance for a shares-based Clause? If so, what steps should drafters take to maximise the likelihood that the shares are considered ‘unique’? What other issues might result from a shares-based Clause?
What options are available to reduce the risk that a third-party acting in good faith and paying valuable consideration acquires irreversible IP rights in pre-WGAI? Do any common law jurisdictions allow for stronger property rights which would survive such a transfer? How else could the Developer’s behaviour be circumscribed with contractual prohibitions to further reduce this risk? Could widespread publication of the Developer’s adoption of the Clause encourage the court to reach the conclusion that a third party was acting in bad faith?
What steps can drafters take to lighten the evidential burden on the Counterparty in the event of a breach? If the Developer permitted certain kinds of monitoring or agreed to a regular system of disclosure, could this make it easier for the Counterparty to prove a breach? Could the Counterparty and Developer pre-agree certain behaviours which would automatically entitle the Counterparty to sue for breach?

The Law:

The most common remedy for breach of contract in English law is expectation damages, which is available as of right to an injured party.^[4] The second-most-common remedy is specific performance, which will be available at the discretion of the court. I’m going to consider each remedy in turn before demonstrating why they present a problem for the enforceability of the Clause and considering possible solutions, linking back to the above recommendations as I go.

Damages:

Which losses are recoverable?

When English courts calculate damages for breach of contract, the aim is to place the injured party in the position that it would have been in, had the contract been correctly performed.^[5] This does not imply that all losses^[6] are recoverable - the injured party is only entitled to recover damages for losses that both parties knew, at the time of contracting, might result from a breach.^[7] Additionally, such losses must be more than foreseeable at the time of contracting - they should be ‘not unlikely’.^[8]

How are losses calculated?

The general rule is that the size of a loss is assessed by reference to the value of the thing lost at the date of the breach.^[9]^[10] Where losses are hypothetical the court will estimate the ‘level of the loss’ by evaluating the ‘chances’ of any given loss occurring, unless those chances are no more than ‘remote speculation’ in which case they are discounted entirely.^[11]

Specific Performance:

The court will grant an order for specific performance when it considers damages insufficient to account for losses which have been or could be sustained by the injured party.^[12] It is most frequently ordered where the consideration offered by the offending party is unique or has no market substitute.^[13] However, specific performance is a discretionary remedy which the court will not always provide, even where it considers damages to be inadequate. Most importantly for our purposes, the court will not order specific performance if:

this would force the defendant to breach a contract with a third party, provided that the third party acted in good faith,^[14] or
the defendant and claimant are parties to a contract by deed which is unsupported by consideration.^[15]

How does this affect the viability of the Clause?

The above presents two issues for the Windfall Clause. Firstly, if the Developer breaches the Agreement, standard contractual remedies may be insufficient to make up for the Counterparty’s loss. Secondly, the Developer’s ability to breach the terms of the Agreement without serious financial consequences means it may be insufficiently disincentivised from doing so.

Inadequate remedies:

Remedies for the Developer’s breach are likely to be inadequate for two reasons, which I will address separately:

Poor calibration, remote speculation and the evidential burden:

The first issue is that judges’ beliefs about which losses are ‘not unlikely’ and which are ‘remote speculation’ will become increasingly poorly calibrated as Developers approach WGAI. Particularly if one predicts a rapid acceleration in AI capabilities over time, submissions in court like ‘DeepMind is about to control >1% of gross world product’ might appear outlandish right up until they actually happen, and judges with a sceptical stance on AI development may disbelieve even the strongest arguments about the possibility of a windfall.^[16] Quite apart from the scepticism of the judiciary, the burden of proof for this extraordinary claim will be substantial - to succeed in its claim, the Counterparty will need to prove a highly technical and inherently speculative argument on the balance of probabilities. Furthermore, successfully arguing its case will require that the Counterparty has high-level access to the Developer’s communications records and technical information, which it may take deliberate steps to obfuscate.^[17] This presents a significant risk that English courts will find the odds of a windfall to be ‘remote speculation’ and so unrecoverable, severely limiting the Counterparty’s damages.^[18] These issues will be exacerbated where the relevant breach is breach of a duty to act in good faith as the Counterparty will need to demonstrate not just that a given piece of software was (pre-)WGAI, but that the Developer and any relevant third party knew or should have known this.^[19]

This risk of sceptical or poorly-calibrated judges motivates recommendation 1. To avoid this problem, the parties might decide to insert an expert determination provision in the Agreement, mandating that any dispute as to the quantity of expectation damages available for breach any term be left to a panel of experts chosen by the parties. Most likely, these would be a collection of well-calibrated experts in AI, economics and law, who could provide a more reliable estimate as to the likelihood that the Developer would develop WGAI, and thus the appropriate level of damages for a breach. Expert determinations are rarely set aside by the court,^[20] meaning that the Developer would almost certainly be bound by any such decision, potentially allowing the Counterparty to avoid the issue of poorly calibrated judges entirely. This may also help to counter the issues with establishing bad faith, as the parties could refer any dispute over the Agreement to the expert panel without the need to establish any fault on the Developer's part.^[21]

This risk also motivates recommendation 2. In theory, the insertion of a basis clause stating that the parties agree that the Developer’s development of WGAI is not remote speculation would create a contractual estoppel,^[22] preventing the Developer from claiming the opposite in court.^[23] If the court were willing to uphold the provision, this would ensure that the Counterparty’s loss of possible future distributions of windfall profits would not be considered too remote to be recoverable, which would drastically increase the scale of the damages available to the Counterparty. As with recommendation 1., it is hoped that this would ensure that any estimate of damages by a judge would more closely match the Counterparty’s true loss in expectation.

Note that both of the above recommendations may not be feasible, for reasons I discuss below. However, if they are, they could substantially resolve the issue of unsatisfactory damages.

Sidestepping Specific Performance:

The issue with specific performance as a remedy for breach of the terms of the Agreement is not that it would be unsatisfactory if it were available. Rather, I am concerned that specific performance will be unavailable for many plausible situations in which the Developer defaults on the Agreement. Most importantly, for our purposes:

Specific performance won’t be available if the Agreement is donative - If the Counterparty provides no consideration for the Agreement, it will be unable to bring a claim for specific performance against a Developer.^[24] This robs a purely donative Agreement of a vital remedy for breach and is a strong argument against a one-sided contract.
Specific performance (and other equitable remedies) won’t be available if damages are considered adequate - Seeing as the Clause is a promise to pay money, the courts may consider damages to be an adequate remedy and so refuse to order specific performance of the contract in the event of a breach. As I’ve noted above, any award of damages may be insufficient, meaning that the Counterparty will not receive a satisfactory remedy for the breach.
The court won’t force the Developer to break a contract with a third party - A plausible failure mode of the Agreement involves a Developer selling or licensing its IP rights in pre-WGAI to a third party which is not bound by the Clause in return for a lump sum.^[25] Provided the third party is not acting in bad faith, the court will not enforce the terms of the Agreement by an order for specific performance, meaning the Counterparty might only be able to claim damages. This particular example highlights a more general failure mode of the Agreement as concerns equitable remedies: namely, in any situation where the Developer contracts with a third-party acting in good faith in such a way as to sidestep its obligations under the Clause, the breaching act itself will foreclose the option of seeking equitable remedies for the breach.^[26]

These concerns motivate recommendations 3.-5. in this section, all of which increase the likelihood that specific performance is available in at least some breach scenarios:

Recommendation 3. ensures that consideration moves from the Counterparty, making specific performance available at least in principle.^[27]
Recommendation 4. might encourage the court to provide specific performance in the event of the Developer’s breach. Even if it wasn’t strictly enforceable, the drafters might wish to leave the provision in the Agreement as a bluff clause. I’ve considered this in further in in Appendix III.
Recommendation 5. stems from the fact that English courts consider (some types of) company shares ‘unique’, meaning that they will not consider damages an adequate remedy for breach of contract.^[28] A shares-based Clause could thus increase the likelihood that the court ordered the Developer to perform its continuing contractual obligations rather than pay damages.

Efficient breach:

The recommendations I’ve outlined above increase the likelihood that adequate remedies will be available for the Developer’s breach of contract, but they fail to substantially eliminate the possibility that the Developer will have an incentive to default. For one, they fail to address the unavailability of equitable remedies in the event of good faith transactions with third parties, meaning that a Developer can cash out with relative impunity by selling or exclusively licensing its (pre-)WGAI. Furthermore, it’s unclear whether a court will uphold a basis or expert determination clause, making it unlikely that the Counterparty will receive proportionate damages in the event of a breach. I consider both of these issues in more detail later in this post. For the moment, it’s sufficient to recognise that these shortcomings leave open the possibility that a Developer will deliberately breach the Agreement simply because doing so is in the corporation’s best financial interests.

To formalise this issue slightly - the problem is that whilst remedies for breach of the Agreement remain toothless, a Developer will almost always be able to engage in ‘efficient breach’.^[29] This is because a Developer on the verge of creating WGAI will be faced with two choices:

Perform the contract and definitely pay 10% of all future profits above a certain level to the Counterparty.
Breach the contract and accept an X% chance of paying ≤10% of the relevant profits to the Counterparty plus legal fees, and a (100-X)% chance of paying nothing.

For most values of X, a rational Developer will prefer (b).^[30] This is a serious problem for the viability of an English Windfall Clause because it means that the Developer will usually be incentivised to default on the Agreement.

The importance of reducing the likelihood that the Developer can commit efficient breach provides further justification for both recommendations 1. and 2., as both of these increase expected costs for the Developer if it defaults on the Agreement by increasing the level of damages which a Counterparty could expect to recover. Recommendations 3.-5. have a similar effect because they increase the likelihood that the court will order specific performance which, in turn, increases the odds that a Developer will be forced to pay out under the Clause even if it tries to avoid doing so.

Outstanding issues:

Although the recommendations I have outlined go some way to resolving the issues with remedies, several outstanding problems remain. In the rest of this post, I’m going to highlight and discuss a few issues which appear particularly important as concerns the enforceability of the Clause.

Issue 1 - A basis clause might be considered a penalty:

Under English law, a penalty clause is one which ‘imposes a detriment on the contract-breaker out of all proportion to any legitimate interest of the innocent party in the enforcement’.^[31] If the provision imposes an ‘extravagant’ or ‘unconscionable’ penalty on the offending party for breach of contract, it will be struck out and the court will revert to a standard approach for evaluating damages. Based on this definition, it seems plausible that the basis clause which I have described above will be construed as a penalty. For this reason, I believe it is more likely than not that a basis clause would be unenforceable.^[32]

This requires some unpacking. First, it is important to remember that an injured party can recover for any losses which are not too remote. Second, we must focus on the intended effect of the basis clause - namely, the effect is that the Developer is estopped from claiming that the speculative loss of distributions from future windfall profits are not considered too remote. The practical result of these two facts is that a Counterparty will be able to claim truly colossal damages in the event of a breach, equal to some portion of the value of all the Developer’s future windfall profits. Viewed from the perspective of a sceptical court, it’s easy to see how this could be problematic: a clause requiring one party to pay another billions of pounds for breach of a contract would almost certainly appear ‘extravagant’ or ‘unconscionable’ given that the Counterparty would most likely receive nothing even if the Developer performed the contract.^[33] The fact that the Counterparty had not provided commensurate compensation for the Developer’s promise could only underscore this impression, as such a payment would seem to be wildly out of proportion to any legitimate interest that the Counterparty might have in the enforcement of the contract.^[34]

Should the court find that the basis clause was a penalty it would be struck out, leaving the Counterparty to rely on standard contractual damages which, as we have seen, would provide wholly inadequate compensation. The likelihood of such a ruling thus suggests that the drafters of an Agreement should be hesitant to rely on a basis clause alone to ensure that the Counterparty can enforce the Clause.^[35]

Issue 2 - An expert determination clause may not provide adequate remedies:

As I’ve noted above, one possible response to the issue of poorly-calibrated judges is to refer disputes over damages for breaching any term of the Agreement to a panel of experts who could provide with a more accurate forecast of the Counterparty’s loss in expectation. Unfortunately, I can foresee three potential problem with this approach:

A Developer would argue for appointing AI sceptics - A Developer is unlikely to agree to a full panel of longtermists with short-timelines. To the contrary, it will be incentivised to appoint experts who are sceptical about the progress of AI and thus more likely to provide conservative estimates that it will develop WGAI. Although it seems reasonable to have a diversity of voices on any panel, the effect of this is that the panel’s calculation of damages could systematically skew too low. This would mean both that the Counterparty would not receive satisfactory damages in the event of a breach and that the Developer would be incentivised to engage in efficient breach.
Expert uncertainty about AI progress could lead to unsatisfactory decisions - There is significant disagreement between experts on the rate and timing of AI progress, suggesting that a panel’s estimates could vary wildly depending on the experts appointed.^[36] This could be worsened by the fact that it will be difficult to predict in advance which experts will be well-calibrated at the time a Developer breaches its obligations.^[37] The result of this is that any damages awarded by such a panel may not closely track the probability that the Developer will develop WGAI.
Information asymmetries could encourage the Developer to engage in efficient breach - This issue is closely related to the issue above. Essentially, a Developer will know more than any outside observer about its likelihood of developing WGAI at any given time. Where a Developer believes that an expert panel’s predictions about AI timelines are short of the mark, it may thus be incentivised to commit efficient breach.^[38]

Given the difficulty of answering questions around AI timelines, decisionmakers will undoubtedly remain divided, meaning that any assessment of damages will depend heavily on the intuitions of whoever makes the decision. Therefore, these problems are not related to expert determination per se so much as they are reflective of the issue with relying on damages as a remedy for the Developer’s breach. Nonetheless, they provide good reason for drafters of an Agreement not to rely entirely on expert determination and to seek additional methods of enforcing the terms of the Agreement.

Issue 3 - There is no obvious way to stop a third party from acquiring rights in the relevant IP:

I’ve already covered this issue in detail above, so I won’t rehash it here. However, I want to emphasise that this is a very serious issue for the viability of the Clause, because it grants the Developer and its key stakeholders an opportunity to irreversibly sidestep their obligations whilst continuing to extract value from the development of WGAI. In the interests of making this problem more salient for readers, I’m going to give an example of how a Developer could use the third party loophole to break its promises in spirit, though not in law:^[39]

Suppose that a sophisticated majority shareholder in the Developer believes that one its business units has a high chance (say, 80%) of developing a product that will trigger the Clause in the near term, but that belief is not widely shared or known. That shareholder might try to force A to sell the business unit to a second corporation, B (which does not have a Windfall Clause) in exchange for shares in B, with the shares in B distributed to A's shareholders on a pro rata basis as a dividend.

The Counterparty could argue that the sale is a breach of a term in the Agreement to perform the contract in good faith. This is almost certainly the case. However, to receive an order for specific performance for that breach, the Counterparty would have the burden of proving:

the value of the business unit as of the time of the transfer;
that the value was high enough to trigger the Clause;
that the Developer deliberately or negligently breached its obligations under the Agreement; and,
that the third party purchased the (pre-)WGAI in bad faith.

In situations such as this, a Counterparty will already be in the unenviable position of having to convince the court that a single business unit will soon be worth >1% of GDP, and that the Developer was acting in bad faith as it knew or should have known this. By introducing corporation B into the equation, the Counterparty is now saddled with a further requirement of proving that the third party was acting in bad faith - for example, by trying to help the Developer sidestep its obligations under the Clause. Notwithstanding that the third party might, in fact, be purchasing pre-WGAI unwittingly and without bad intent, the evidential burden will be difficult for the Counterparty to discharge even if B was acting in bad faith.

I will not consider this example further, but I find it very troubling as concerns the viability of the Clause. I predict that this is not the only example of methods whereby a Developer receiving good legal advice could obviate the need to pay out under the Clause by transferring pre-WGAI to a third party, and if this is so it might be near-impossible to prevent a determined Developer from breaching the contract.

Issue 4 - The evidential burden to be discharged by the Counterparty remains high:

As noted above, the Counterparty will only succeed in any claim if it can discharge a substantial evidential burden, convincingly laying out a speculative and highly technical argument about the likelihood that the Developer in breach has or had possession of (pre-)WGAI, which could make it almost impossibly difficult to recover for any breach. Quite apart from any inherent scepticism by the judges, this is simply a hard case to make. This issue might be ameliorated by imposing ongoing disclosure requirements on the Developer, allowing the Counterparty to conduct regular evaluations of whether a given technology is or is not WGAI. However, it is doubtful that a Developer would consent to the extensive external monitoring which would be required for this approach to succeed. In the absence of such an approach, it’s unclear how the Counterparty could lighten its evidential burden enough to reach an acceptable probability of success in court.

Conclusion - The (English) Windfall Clause (probably) has a remedies problem:

The viability of the Windfall Clause hinges on the adequacy of its remedies, and the remedies currently available are unfit for purpose. I have shown that it is possible to improve this situation, but I don’t think any of the solutions I have proposed bring the Clause to a point where it could not be easily sidestepped by a well-advised Developer intent on getting out of its obligations. This is a serious problem for the English formulation of the Clause and, given that the general principles for calculating damages are mostly uniform across common law jurisdictions, this could be a serious problem for the Clause overall. Taking all this into consideration, I believe EAs interested in policymaking should significantly decreased their confidence in the Windfall Clause as an example of good longtermist policy.

Before concluding, I should stress that there is a good chance I am making mistakes here. Being a law student, I lack the knowledge and experience to reach firm conclusions on this topic. This motivates recommendation 6., as I believe qualified Counsel would be far better equipped to consider the issues I have raised and suggest effective solutions. I also invite any interested legal professionals reading this to continue my investigation. By discussing this issue before the Clause is widely adopted, I hope to ensure that we implement a Windfall Clause that really works and, if this is impossible, that we continue our search for effective longtermist policy recommendations elsewhere.

Appendices:

Appendix I - Wouldn’t trading away the IP in pre-WGAI be a transaction defrauding creditors?

In English law, a Developer knowingly trading away pre-WGAI might be considered a transaction defrauding creditors (‘TDC’), allowing the court to unwind the transaction so long as it is satisfied that the transaction was entered into for the purpose of ‘prejudicing the interests’ of the Counterparty.^[40]^[41] Unfortunately, there are two important limits to the powers granted by the Insolvency Act which mean it will not prevent a determined Developer from sidestepping its obligations under the Agreement by transfers to a third party:

TDC will only occur where the Developer has received ‘significantly less’ consideration from the third party than it has given. As I’ve noted repeatedly in the body of this post, I expect courts will be highly sceptical of claims that the Developer is on the verge of controlling a significant portion of global GDP, particularly if the Developer is strenuously denying the same. This will make it difficult, perhaps impossible, to demonstrate that a third party purchaser of pre-WGAI is paying disproportionately little consideration.
Even if the Counterparty can demonstrate that the third party has provided inadequate consideration, this will not foreclose all plausible options for the Developer to sidestep its obligations. This is because the Insolvency Act prohibits the court from prejudicing any third party’s interest in the property which was fraudulently transferred, provided that interest was acquired in good faith and without notice from a person other than the debtor.^[42] Some inventive avoidance methods (such as the one outlined in Issue 3 above) allow the shareholders of a Developer to extract value from WGAI without transacting with the Developer. Provided the shareholders responsible for the transfer could plausibly deny in court that they believed the relevant AI system was pre-WGAI at the time of the transaction, the Counterparty could not force the Developer to unwind the transaction and would only be entitled to damages.^[43]

Finally, note that the Counterparty will only be able to bring a TDC action if it can demonstrate that it is a creditor in the first place. As we have seen, the Counterparty may struggle to discharge the evidential burden needed to succeed in its claim, in which case it will not be entitled to any remedy.

Appendix II - Couldn’t the Counterparty just apply for an injunction to stop the Developer from breaching the contract in the first place?

Prohibitory injunctions will be ineffective to prevent the Developer from breaching the contract in most plausible default scenarios because a Developer is unlikely to publicly announce that it intends to breach the Agreement. In fact, a Developer might breach a term of the Agreement unknowingly; for instance, uncertainty about the potential of a particular AI system might lead a Developer to sell (pre-)WGAI without realising it. However, once the Developer has decided to act in such a way as breach a term of the Agreement, it could carry out the breach faster than the Counterparty could detect - note, for example, that IP rights in AI systems can be transferred instantaneously - making it impossible for the Counterparty to apply for a prohibitory injunction in time.^[44]

Appendix III - Would the court uphold a specific performance clause?

I think it’s unlikely that the court would uphold a specific performance clause because, as a general principle, English courts are hostile to attempts to fetter their discretion to provide equitable remedies.^[45] I haven’t been able to find any judgments in the higher courts which consider this question, so there’s an outside chance that the provision would succeed, though a widely-cited source of commentary indicates that courts would not uphold such a clause.^[46]

Even if this is true, there is reason to include the clause in the Agreement. Firstly, I expect it would lightly encourage the court to provide specific performance where the judge is uncertain about the adequacy of damages. Secondly, even an unenforceable provision could operate as a bluff clause, discouraging the Developer from breaching the contract in the first place.

Appendix IV - Could the parties place a trust over WGAI?

Given the issues with recovery in the event of transfer to a third party, it might be desirable to provide the Counterparty with a beneficial interest in any pre-WGAI or WGAI itself, such that it could follow the IP into the hands of the third party. One option which might achieve this goal is to draft the Agreement so as to automatically assign some share of beneficial title in pre-WGAI to the Counterparty at the moment it is developed.

One issue I can foresee with this is that it would be very difficult to overcome the evidential burden required to show that any given AI system is WGAI. Also, I should highlight that I haven’t explored this in any detail and my knowledge of trust law is very rusty. I’ve included it nonetheless to help others interested in exploring this solution avoid duplicated work.

^{^}
Thanks to Cullen O’Keefe for this recommendation. A ‘basis clause’ is a term used by insurers to describe a clause which converts a pre-contractual representation by the insured into a warranty. This isn’t strictly what I’m describing here, but it’s the closest analogue I can find for the concept I’m trying to outline. If you're interested in exploring this topic further, another useful search term is 'non-reliance clause'.
^{^}
Alternatively, the drafters might insert a provision stating that equitable remedies will be provided irrespective of whether the Counterparty has provided consideration. English courts might not uphold this alternative provision for much the same reasons as I outline in Appendix III, but it may still be worth including as a bluff clause.
^{^}
I’m planning to consider this question throughout the sequence, but I don’t consider it again in this post. Watch this space.
^{^}
Note that I don’t consider the possibility of a debt action by the Counterparty in this post, although this is another common remedy for breach of contract. This is because the likelihood and amount of any distribution under the Windfall Clause will always be too speculative to allow for a debt action. I have also ignored restitution damages as these would be clearly inadequate to compensate for any breach.
^{^}
Robinson v Harman (1848) 1 Ex Rep 850, Parke B at 855
^{^}
When I talk about ‘losses’ please assume that I am also discussing foregone gains such as the Counterparty’s chance of receiving a share of windfall profits.
^{^}
Hadley & Anor v Baxendale & Ors [1854] EWHC J70
^{^}
Koufos v C Czarnikow Ltd (The Heron II) [1967] UKHL 4.
^{^}
The general principles I outline here applies to all losses in expectation. However, in this section I focus specifically on the Counterparty’s loss of future distributions from windfall profits which have yet to be made at the time of the Developer’s breach. This is for two reasons. Firstly, because I don’t foresee any issues with remoteness should the Counterparty sue the Developer for windfall distributions which are already in arrears. Secondly, because most plausible failure modes for the Agreement involve a breach by the Developer before or shortly after its obligations under the Clause are triggered. This means that most of the damage suffered by the Counterparty is likely to be the loss of hypothetical distributions, making it particularly important that such losses are recoverable.
^{^}
Miliangos v George Frank [1976] AC 443; Dominion Mosaics Limited v Trafalgar Trucking Co Limited [1990] 2 All ER 246; and Golden Strait v Nippon Yusen (the ‘Golden Victory’) [2007] UKHL 12. However, note also that damages may be calculated at a later date where the claimant does not treat the contract as repudiated and first seeks specific performance: Rahman v Rahman & Ors [2020] EWHC 2392 (Ch). This mildly weakens my argument in this section - if the Developer breached the contract prior to developing WGAI but the Counterparty’s claim was only settled afterwards, the Counterparty might receive far more in damages.
^{^}
The language here is frustratingly imprecise, making it difficult to give a more accurate summary than I have given above. See Parabola Investments Ltd & Ors v Browallia Cal Ltd & Ors [2010] EWCA Civ 486 at [23]: ‘Where [calculating loss] involves a hypothetical exercise, the court does not apply the same balance of probability approach as it would to the proof of past facts. Rather, it estimates the loss by making the best attempt it can to evaluate the chances, great or small (unless those chances amount to no more than remote speculation), taking all significant factors into account…’. See also Vasiliou v Hajigeorgiou [2010] EWCA Civ 1475 at [25]: ‘Where the quantification of loss depends upon an assessment of events which did not happen the judge is left to assess the chances of the alternative scenario he is presented with. This has nothing to do with loss of chance as such. It is simply the judge making a realistic and reasoned assessment of a variety of circumstances in order to determine what the level of loss has been.’
^{^}
Rainbow Estates Ltd v Tokenhold Ltd [1999] Ch 64, at paragraph 73.
^{^}
For example, it is common for courts to order specific performance for contracts concerning interests in land, the sale of shares in private companies, and particularly unusual goods. See Verrall v Great Yarmouth Borough Council [1981] QB 202; Gaetano Ltd v Obertor Ltd [2009] EWHC 2653 (Ch); MSAS Global Logistics Ltd v Power Packaging Inc [2003] EWHC 1393 (Ch); and Falcke v Gray (1859) 4 Drew 651.
^{^}
Warmington and another v Miller [1972] 2 All ER 372.
^{^}
Jefferys v Jefferys (1841) Cr & Ph 138. This results from the equitable doctrine that ‘equity will not assist a volunteer’.
^{^}
As an indicator of this problem, consider the fact that lots of very smart people are sceptical of the risks posed by superintelligence. I don’t see a good reason to believe that WGAI will produce different intuitions. To exacerbate the issue, even judges who believe in the risks from advanced AI may fall prey to a host of cognitive biases causing them to round small probabilities of a Developer creating WGAI down to zero and so considering them too remote to be recoverable.
^{^}
Consider, for instance, that key stakeholders of a Developer who force it to renege on its obligations could hold their pivotal conversations away from the workplace, unrecorded and in private.
^{^}
Note that this envisions a scenario in which the Developer defaults on the Agreement before it reaches windfall profits. A court is unlikely to be sceptical if the Developer has, in fact, made such profits.
^{^}
Of course, not all breaches of bad faith duty will involve this particular question, and some might present a less onerous burden to discharge.
^{^}
There’s no single authority for all the circumstances in which an expert determination will be overruled, but I’m confident none of them apply here. If you’d like to find out more, here is a helpful practice note on Practical Law discussing the topic. See also Jones v Murrell [2016] EWHC 3036 (QB) for recent consideration of when the court will overrule an expert.
^{^}
The obvious downside to doing this is that a Developer will want to reserve the right to deal with its IP in good faith. As such, it is unlikely to accept a contract term which stipulates that behaviours which do not actually constitute a breach of contract are referred to expert determination.
^{^}
For example, the parties might insert something to the effect of: “The parties agree that it is not remote speculation that [the Developer] will reach a level of profits which triggers the Windfall Clause and that it is realistic and reasonable to expect that [the Developer] will reach such a level. The parties contract on this basis and may not assert otherwise before any court of law.”
^{^}
The key case on contractual estoppel and basis clauses in English law is Peekay Intermark Ltd v Australia & New Zealand Banking Group Ltd [2006] EWCA Civ 386. This is a relatively new and developing doctrine which has mostly, thus far, been used for non-reliance clauses to limit a party’s liability for misrepresentation. It remains to be seen whether contractual estoppel could be extended to the Windfall Clause, though there is no obvious reason why it could not.
^{^}
Note also that the court may decline to award specific performance if consideration is inadequate and the claimant has acted inequitably. For example, see obiter in Falcke v Gray (1859) 4 Drew 651 at [659]: ‘The general rule with regard to hard bargains is that the Court will not decree specific performance, because specific performance is in the discretion of the Court for the advancement of justice.’ If the Counterparty acted in a way which a court deemed inequitable - for example, by delaying before it brought a claim against the Developer - this might foreclose the option of specific performance. I’m quite unsure about this claim and haven’t looked into the topic in detail.
^{^}
See O'Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J. and Dafoe, A., 2020. The Windfall Clause - Distributing the Benefits of AI for the Common Good. [online] at pp.6-7: The authors of the original report suggest that this is unlikely, because it would not be in the Developer’s long-term best interest to do so. I agree that it might not be in the Developer’s best interests, but I don’t think this entails that a Developer won’t trade away (pre-)WGAI. For one, businesses typically discount future earnings. It seems very plausible to me that a purchaser with a lower discount rate, such as a national government, might seek to buy WGAI from the Developer for a very large lump sum. Second, it seems unreasonable to expect a Developer to be totally rational in the event of windfall profits. Group preferences are (usually) intransitive, meaning that the Developer may sell (potentially) WGAI even if this is harmful to the business. Furthermore, shareholder and managerial short-termism are commonplace in large firms, and key stakeholders might decide to cash out by forcing the company to sell WGAI even if this isn’t in the best interests of the Developer as a whole. This may be further encouraged by scope neglect and the diminishing marginal utility of money for individual stakeholders. See the efficient breach section and note 30 below for further discussion of why a Developer might trade away (pre-)WGAI even if it is in the Developer's best interests to refrain from doing so.
^{^}
Thankfully, the court will order specific performance where the third party is a sham company set up for the purposes of evading specific performance: see Jones and another v Lipman and another [1962] 1 WLR 832 and Prest v Petrodel Resources Ltd & Ors [2013] UKSC 34. This helps avoid one obvious failure mode of the Agreement, in which the Developer sets up a sham company with the same shareholders to which it then transfers WGAI. The court may also reverse the transaction if it is a ‘transaction defrauding creditors (see Appendix I).
^{^}
Cf. the Founder’s Pledge contract which, perhaps for this reason, is not framed as purely donative.
^{^}
See supra note 13. Note that if the Developer is a publicly traded company it is essential that the drafters stipulate unique properties for the shares, such that they could not otherwise be acquired on the market.
^{^}
Efficient breach is a term in law and economics which describes a situation where it is cheaper for a party to breach a contract and pay damages than it is for that party to perform the contract.
^{^}
I’m just trying to give a sense of the general character of the problem here, rather than provide a high-fidelity game theoretic model. As such, I recognise that I haven’t accounted for less tangible economic benefits which could accrue to the Developer, such as positive publicity, the reduction of political risk and improved employee goodwill. I also haven’t considered that the Developer will be governed by shareholders and high-ranking employees whose intrinsic pro-social motivations may further reduce the true expected value of a decision to breach. I’ve noted in the takeaways section that this is an issue for further investigation, but these aren’t legal questions so I won’t consider them any further here. Nonetheless, your beliefs about these issues should influence the values of X for which you think a Developer can engage in efficient breach and thus your overall impression of the viability of the Clause.
^{^}
For the most recent statement of the law on penalties, see Cavendish Square Holdings BV v Makdessi; ParkingEye Ltd v Beavis [2015] UKSC 67. Note that the true test for a penalty clause is more complicated than I have outlined in the body of this post - see here for a helpful summary. Also, note that Cavendish partially overturned the ruling in Dunlop Pneumatic Tyre Co Ltd v New Garage & Motor Co Ltd [1914] UKHL 1 (1 July 1914), which states that a clause is a penalty where it is not a ‘genuine pre-estimate of loss’. If Dunlop is still good law in the other jurisdictions I consider in this sequence, this could be highly problematic for a basis clause outside of England and Wales. Finally, note that most of the case law around penalties relates to liquidated damages provisions, whereas a basis clause concerns an unspecified sum and might be treated differently by courts as a result.
^{^}
I'm not confident in this belief as I've not researched this topic in great depth, so I could be missing something here.
^{^}
Note that the income which the injured party expected to receive prior to the offending party’s breach of contract is important in determining whether a clause is penal. For example, see Unaoil Ltd v Leighton Offshore Pte Ltd [2014] EWHC 2965. Note however that Unaoil involves a liquidated damages provision, meaning that the Dunlop test for a ‘genuine pre-estimate of loss’ is still relevant to determining if a clause is penal.
^{^}
Note that, as Cavendish is a recent judgment, ‘legitimate interest’ remains poorly defined. However, as an indication, case law prior to Cavendish mostly considers whether the injured party has a justifiable commercial interests in the enforcement of a purported penalty clause, such as an interest in avoiding delay (Azimut-Benetti SpA (Benetti Division) v. Darrell Marcus Healey [2010] EWHC 2234 (Comm)) or as reassurance against premature termination of employment (Murray v Leisureplay plc [2005] EWCA Civ 963). With this in mind, any case brought by the Counterparty would need to emphasise that it has a commercial interest in enforcing compliance with the Agreement, perhaps by highlighting that a Developer’s successful breach might create a domino effect of breaches across the industry. Also note that there is a presumption that in a negotiated contract between properly advised parties of comparable bargaining power, the parties themselves are the best judges of what is a legitimate interest (see, for example, De Havilland of Canada Limited v SpiceJet Limited [2021] EWHC 362). This could aid the Counterparty in court, allowing it to argue that the very nature of the Agreement involves both parties agreeing that the Counterparty has a legitimate interest in binding the Developer to its promise.
^{^}
Note that this doesn’t mean that the drafters shouldn’t include a basis clause, just that it might not work. In the worst-case scenario, a basis clause could be an effective bluff, discouraging the Developer from default in case the clause were effective, without costing the Counterparty anything to enforce.
^{^}
Scott Alexander’s Yudkowsky Contra Christiano On AI Takeoff Speeds is a good illustration of just how much even well-calibrated experts disagree on fundamental questions concerning AI development. I see no reason why these same sorts of disagreements wouldn’t occur here.
^{^}
This issue cannot be resolved by appointing the experts after a Developer has breached a term of the Agreement because at this point the Developer will be even more incentivised to appoint AI sceptics.
^{^}
It might be possible to resolve this by giving experts a high level of continuous access to a Developer’s confidential information, but it seems unlikely that a Developer would be happy to submit to this sort of intrusion on an ongoing basis. I discuss this briefly later in the post.
^{^}
Thanks to Will Greenman, whose comment I have heavily paraphrased here, for inspiration in creating this example.
^{^}
Sections 423(3) and 425(1) Insolvency Act 1986 (UK). Note that the rules around TDCs are more complex than I’ve outlined here.
^{^}
This is the least-researched part of the entire post and may contain factual errors. If this doesn’t seem correct, let me know in the comments. Thanks to Cullen O’Keefe for raising this issue.
^{^}
Section 425(2) Insolvency Act 1986 (UK).
^{^}
This is a weaker argument than my first suggestion, because it relies on the additional supposition that a sophisticated stakeholder in the Developer will form a detailed plan to deceive the court.
^{^}
Injunctions are not available as of right, but they are almost always available to prevent a breach of contract. For a helpful discussion of the circumstances in which English courts will grant an injunction, see Contractual Duties: Performance, Breach, Termination and Remedies (Commercial Series, 3rd Edn Sweet & Maxwell, 30 Jul 2020) para 28-008. See also Doherty v Allman (1878) 3 App Cas 708 at 720; Insurance Co v Lloyd’s Syndicate [1995] 1 Lloyd’s Rep. 272 at 277; Priyanka Shipping Ltd v Glory Bulk Carriers Pte Ltd [2019] 1 W.L.R. 6677 at 97. For a helpful statement of the principles a court will apply when considering granting an injunction to prevent breach of a negative covenant, see SDI Retails Services Ltd v Rangers Football Club Ltd [2018] EWHC 2772 (Comm) at paragraphs 43, 47-51. For a case which is (mildly) relevant to the Windfall Clause, see also Donnell v Bennett (1883) L.R. 22 Ch. D. 835: in this case, the court granted an injunction to prevent sellers from selling their output to anyone other than one buyer.
^{^}
For example, see Quadrant Visual Communications v Hutchison Telephone (UK) [1991] 11 WLUK 378. Thanks to Peter Wills for highlighting this issue.
^{^}
Snell’s Equity (Trusts, Wills and Probate, 34th Edn Sweet & Maxwell, 13 Dec 2021) para 17-004: ‘The court’s discretion to order specific performance cannot be fettered by a term in the contract purporting to oust the remedy, because to give effect to such a term would reduce the court’s function “to that of a rubber stamp”, and it seems likely that the same would apply to a term purporting to require specific performance.’ Thanks to Peter Wills for highlighting this issue.

Effective Altruism Forum
EA Forum