Hide table of contents

Over the coming months, I will be posting the results of a series of shallow investigations into the viability of the Windfall Clause in key jurisdictions for the purposes of AI Governance. If you're not already familiar with the Windfall Clause, this talk by Cullen O'Keefe does a great job of explaining how it works.

The goal of this sequence is twofold. First, I’d like to highlight key considerations for non-US EA lawyers when drafting agreements containing a Windfall Clause. I hope this will lay the groundwork for future efforts to put the Windfall Clause into practice at the world's top AI developers. Second, I want to make sure that policy advocacy for a Windfall Clause is based on a good understanding of whether it will actually work. If the EA community is to champion the Clause as an example of good longtermist policy, it's essential that it is both legally and politically feasible.

I'm planning to update the sequence as I go, so let me know if you feel I am missing something important and I will update the posts to include them. Alternatively, you are welcome to write a supplement to the sequence discussing anything you feel I've missed. In either case, you can reach me at johnmichaelbridge[at]gmail.com.

A Note on Legalese

This sequence is chiefly written for EAs with legal training, meaning I make extensive use of phrases like 'promissory estoppel' and 'counterparty'. If you are a non-lawyer reading this, I understand you may be put off. If you'd like me to explain anything in the sequence in more detail, please feel free to ask in the comments below or reach out directly.

A Note on Numbers

First thing's first - I apologise for the lack of numbers in this sequence. There are two reasons for this:

  1. I, like many lawyers, totally suck at statistics. This is something I'm trying to fix (tips much appreciated). For that reason, I've decided against trying to give precise probability estimates in this sequence because I think they will decrease the clarity of my writing.
  2. It is not always unhelpful to try to quantify legal terms of art - Judgments from around the world are replete with poorly-defined discussions of probability involving terms like 'a bare possibility', 'beyond reasonable doubt', 'not unlikely' and 'a reasonable chance', and most courts are allergic to maths. The consequence of this is that legal terms of art often mean different things at different times - for instance, a judge stating that something is 'not unlikely' can signal anything from 10-70% confidence in a proposition, depending on the case. Therefore, attempting to quantify these terms might actually decrease the clarity of this sequence.

Defined Terms

Another note: for brevity's sake, and to avoid writing the word 'windfall' even more than I have already, I will be using a few defined terms throughout the sequence. Specifically:

  • The Developer is the AI-developing business, non-profit or government body that promises to pay out under the Windfall Clause.
  • The Counterparty is the recipient of any windfall, whose task is to distribute that windfall for the common good.
  • The Clause is the binding term, or set of terms, defining the Developer's payment obligations to the Counterparty.
  • The Agreement is the document, or set of documents, which determine the full scope of the agreement between the two parties.
  • WGAI, short for 'windfall-generating AI', is the AI system (or collection of systems) capable of generating windfall profits for the Developer. For the sake of brevity, this will also refer to the IP in windfall-generating AI where appropriate.
  • Pre-WGAI is AI which will eventually generate windfall profits, but has yet to do so.
  • TAI is short for 'transformative AI', which is 'AI that precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution'.[1] Pre-TAI is, predictably, AI which will eventually precipitate such a transition.

Which Jurisdictions?

The next seven posts will cover England, India, Israel, Canada, Singapore, the BVIs and Ireland. I have chosen these jurisdictions because they are either:

  • a) locations where transformative AI is likely to be developed, or
  • b) jurisdictions where top AI developers could be domiciled, or
  • c) the sources of law which often govern international agreements like the Clause.

Right now, we don't know where and by whom transformative AI will be developed. As such, we can't say with certainty which law will govern the Agreement if and when a Developer makes windfall profits. However, it's likely that the rules in at least one of these jurisdictions could apply: the first five countries I've listed are some of the world's top investors in AI,[2] the BVIs are an international finance and business hub, and many of the world's largest tech firms have their European HQ in Ireland.[3]

Why no Civil Law jurisdictions?

Having been trained in a common law system, I lack the expertise to do a high-quality investigations into civil law jurisdictions like Germany or France. Nonetheless, I recognise that the EU could be a central player in the development of transformative AI. For that reason, I welcome contributions from any other EA lawyers who are interested in writing a civil law supplement to these posts.

Which Issues?

FHI's original report considers a host of potential legal and political barriers to the successful implementation of the Clause.[4] Considering all of them in detail is beyond the scope of this sequence - I will discuss only the subset of issues which are concerned with questions of law and which present a clear challenge to basic implementation and enforceability of the Clause. These issues can be roughly divided into three parts:

First, before the Agreement is signed, it is crucial to explore any legal restrictions which could prevent the Developer and Counterparty from agreeing to it in the first place. With this in mind, I will consider the following questions for each jurisdiction:

  1. Could signing the Agreement be a breach of a directors' duty to act in the best interests of the company?
  2. What are the legal restrictions on the size and nature of charitable donations, if any?
  3. If the Clause allows the Developer to pay the Counterparty by issuing share options, rather than paying cash, what steps must be taken to create and issue these shares?

Second, after the Agreement is signed, it is important to make sure adequate safeguards are in place such that the Developer cannot sidestep its obligations by legal sleight-of-hand. As such, I ask:

  1. Will a duty of good faith be implied between the parties? If not, can such a duty be expressly introduced?
  2. Can a company pay out dividends to shareholders before meeting its obligations under the clause? If so, is it possible to contract around this?

Third, if the Developer does renege on the agreement, it is vital that the Counterparty has a good cause of action to enforce the Agreement. Therefore, I consider:

  1. If the Agreement is a binding contract, what would constitute good consideration?
  2. If the Agreement is purely donative, could it still be enforced?

Finally, as concerns all three issues above, I will consider if there are there any unique facts about each country’s legal system which could affect the functioning of the Clause.

If you would like to know why I've chosen these specific questions, keep reading. Otherwise, you can skip the remainder of this post and move to later posts in the sequence.

Why these issues?

Could signing the Agreement be a breach of a directors' duty to act in the best interests of the company?

All of the jurisdictions under consideration require directors to act in the way they believe in good faith will promote the success of the company. In Delaware law this is the business judgment rule which, roughly speaking, requires directors to act in a way which they believe maximises shareholder value in expectation.[5] If directors fail to do so, shareholders can apply to the court to block or rescind the transaction or require the directors to account for the loss.[6] 

As the authors of the original report demonstrate, agreeing to the Clause is permitted under the business judgment rule.[7] However, corporate culture, definitions of business success and the specifics of directors' duties vary across jurisdictions. To ensure that the Agreement is not rescinded by disaffected shareholder and that directors are not left liable for breach of duty, it is essential to understand the scope of the best interests duty.

Different jurisdictions have different cultures and legal rules around donations and corporate social responsibility. For example, English company law requires majority shareholder support to make political donations.[8] Meanwhile, Indian law mandates that companies over a certain size donate at least 2% of their profits to charity,[9] and in Delaware, corporate donations are restricted by a requirement of reasonableness to roughly 10% of profits.[10] 

If the Clause is to be purely donative it is thus important to determine the nature and extent of these regulations in each jurisdiction, in case they would prevent a Developer from entering into the Agreement.

Will a duty of good faith be implied between the parties? If not, can such a duty be expressly introduced?

The authors of the original report recognise that a Developer who looks set to develop transformative AI will be incentivised to take any and all steps possible to avoid triggering its obligations under the Agreement.[11] Under US contract law, this risk is mitigated by a duty of good faith implied into business-to-business contracts.[12] This means that if an American Developer tried to dishonestly sidestep its obligations, the Counterparty would have a valid cause of action against the Developer and could sue to enforce the terms of the Agreement.

Given the experimental nature of the Windfall Clause and the unique market conditions which would be created by transformative AI, good faith duties such as these will be key to ensuring that the Developer keeps its side of the bargain. This makes it important to be certain that all of the legal systems under discussion will either imply such duties into the Agreement, or allow their inclusion as express terms.

Can a company pay out dividends to shareholders before meeting its obligations under the clause? If so, is it possible to contract around this?

Under Delaware law, companies are typically required to meet contractual obligations before paying out a dividend to shareholders.[13] This restriction is an important element of an enforceable Clause as it prevents an unscrupulous Developer from avoiding triggering their obligations by paying out large dividends. Determining whether this same restriction applies across jurisdictions is thus essential in establishing the international enforceability of the Clause.

If the Clause allows the Developer to pay the Counterparty by issuing share options, rather than paying cash, what steps must be taken to create and issue these shares?

On one formulation of the Clause, the Developer issues non-voting share options to the Counterparty equivalent to a given percentage of the Developer's profits.[14] A key limitation of this approach in the US context is that amendments of a Delaware corporate charter require shareholder consent, and such amendments are necessary to create and issue a new class of share. This creates an additional veto point in the introduction of the Clause, making a shares-based approach undesirable unless the Agreement has broad company support.[15] 

Shareholder control over share issuance and the company constitution differs considerably worldwide. With that in mind, there is value in exploring which steps would be needed to implement a shares-based approach in the law of each jurisdiction.

If the Agreement is a binding contract, what would constitute good consideration?

Rules around consideration are highly variable across states, and sometimes even within states. For example, no consideration is needed in Scots law, but a Welsh contract cannot be enforced without it.[16] However, the authors of the original report primarily consider the position in the United States, where both parties must give consideration that is sufficient but need not be adequate.[17] This allows the Counterparty to enforce the Clause provided it has given the Developer at least something in return, such as a $1 payment or publicising of the Developer's signatory status.

If the Agreement is to be enforceable as a contract, as in the US, it is thus important to determine whether the Counterparty must give consideration, and what sort of consideration must be given.

If the Agreement is purely donative, could it still be enforced?

Under American contract law, the doctrine of promissory estoppel allows a party to enforce another’s promise even if that party has given no valuable consideration for the promise.[18] Specifically, the defendant may be estopped from breaking their promise if the plaintiff has relied on that promise to their subsequent detriment. This is beneficial to an American Clause, because it allows the Counterparty to enforce the terms of a purely donative Agreement.

However, promissory estoppel does not exist in the same form in every jurisdiction. For instance the English version of the doctrine is famously “a shield, not a sword”,[19] meaning it does not provide an independent cause of action allowing a claimant to enforce a promise. It is therefore important to determine whether a purely donative agreement could be enforced by the Counterparty in every jurisdiction, or whether a binding contract would be needed.

Jurisdiction-specific considerations

Every common law jurisdiction has its roots in English common law. As such, there are striking similarities between all of those I have researched. Nonetheless, some are more similar than others: whereas island nations like the BVIs still rely on the British Privy Council as their final court of appeal, keeping English and BVI law somewhat uniform, India and the United States have diverged considerably since independence.

Divergent features of each legal system could thus present unique challenges in enforcing the Agreement, and I will consider this directly at the end of each post in the sequence.

I've tried to consider only the most pressing legal questions when doing research for this sequence. However, there are undoubtedly further considerations I have neglected to discuss for lack of time. For example, the authors of the original report identify that the Contract may need to include a stipulation that companies in the same group become signatories to the Contract, to prevent intra-group transfers of windfall profit-generating IP.[20] My preliminary investigations suggest that this approach will work across all the above jurisdictions, but I have not looked into it in any great detail.

As noted above, if you feel I am missing key questions like the above, please let me know. I can then either include them in future posts or help you to write a supplement to my posts discussing these issues.

  1. ^

    Karnofsky, H, 2016. Some Background on Our Views Regarding Advanced Artificial Intelligence. [online] Available at: <https://www.openphilanthropy.org/research/some-background-on-our-views-regarding-advanced-artificial-intelligence/> [Accessed 14 November 2022]

  2. ^

    Arnold, Z, 2020. What investment trends reveal about the global AI landscape. [online] Brookings. Available at: <https://www.brookings.edu/techstream/what-investment-trends-reveal-about-the-global-ai-landscape/> [Accessed 7 April 2022].

  3. ^

    The Irish Times. 2022. Top 1000 Irish Companies. [online] Available at: <https://www.top1000.ie/companies> [Accessed 7 April 2022].

  4. ^

    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] pp.15-28. Available at: <https://www.fhi.ox.ac.uk/wp-content/uploads/Windfall-Clause-Report.pdf> [Accessed 7 April 2022].

  5. ^

    Easterbrook, F and Fischel, D, 1996. The Economic Structure of Corporate Law. Cambridge, Mass.: Harvard University Press, pp.98-100.

  6. ^

    For recent judicial consideration of the scope of Delaware courts' discretion to fashion equitable remedies for breach of the BJR, see Basho Technologies Holdco B, LLC v. Georgetown Basho Investors, LLC, C.A. No. 11802-VCL (Del. Ch., July 6, 2018)

  7. ^

    O'Keefe et al. see supra note 4 at 15-16

  8. ^

    Companies Act 2006, s366 (UK).

  9. ^

    Companies Act 2013, s135 (India).

  10. ^

    O'Keefe et al. see supra note 4 at 15

  11. ^

    O'Keefe et al. see supra note 4 at 21-22

  12. ^

    Restatement (Second) of Contracts § 205 (Am. Law Inst. 1981) (US)

  13. ^

    DGCL §§ 170(a)(1), 173. (US/Del)

  14. ^

    O'Keefe et al. see supra note 4 at 25

  15. ^


  16. ^

    Lexisnexis.co.uk. 2022. Key differences in the law of contract between Scots and English law. [online] Available at: <https://www.lexisnexis.co.uk/legal/guidance/key-differences-in-the-law-of-contract-under-scots-law> [Accessed 7 April 2022].

  17. ^

    Restatement (Second) of Contracts § 79(c) (Am. Law Inst. 1981) (US)

  18. ^

    Restatement (Second) of Contracts § 90 (Am. Law Inst. 1981) (US)

  19. ^

    Combe v Combe [1951] 2 KB 215

  20. ^

    O'Keefe et al. see supra note 4 at 22

Sorted by Click to highlight new comments since:

Thanks a ton for doing this, John! Very much looking forward to the rest of the sequence :-)

Thank you for taking on this analysis! Two related issues occurred to me while reading this post. Both fit in your second category of possible "legal sleight-of-hand" by the Developer that could sidestep or dampen the effect of the Clause:

(1) Remedies for breach: What remedies would be enforceable in each jurisdiction for an intentional breach of the Windfall Clause? Assuming courts award compensatory damages instead of ordering specific performance, might they calculate damages based on the expected value of the contract to the Counterparty as of the time of contract formation, rather than the expected value of specific performance as of the time of breach? (If so,  breaching the Agreement upon the triggering event and paying compensatory damages would likely be the Developer's profit-maximizing strategy. Given the amount of value at stake, it seems likely that shareholders would force the Developer to breach the Agreement in this situation, for example by replacing the board of directors.)

(2) Treatment of transactions with shareholders: How would each jurisdiction treat transactions in which a Developer with excess profits agrees to purchase assets from its shareholders at inflated prices? Would the Counterparty have a right of action to block such transactions? (If such transactions are permissible, the Developer could use them to distribute revenue that would otherwise trigger the Clause to shareholders, in lieu of dividends.)

Hi Will,

(1) Is a really good point. I will definitely consider this. A few thoughts right now:

Encouragingly, the position here in England for contractual damages comes from Hadley v Baxendale, where damages are given for all losses that 'were in the contemplation of both parties' at the time they contracted. Given the very nature of the agreement is that the Developer has to pay out a colossal sum if they reach a certain % of GDP/market cap, I'd assume that the Developer's future profits would be included here. That said, specific performance seems like a more satisfactory approach, because I'm sure the court would end up discounting the total sum. 

FWIW, HvB is an old case (1850s I believe), so I expect this same point will apply across most of the former Commonwealth.

(2) I hadn't considered this, but I will  look into it. My initial thoughts are that (a) the shareholders are unlikely to have assets worth >0.1% GDP, so this would be a clear example of bad faith (b) this could be a mark in favour of the stock options method, because it might allow the Counterparty to bring  an unfair prejudice or other derivative action against the Developer.

Thanks for highlighting both.

Hi John,

Thanks so much for the thoughtful reply. On (1), that's great -- I agree that the Hadley rule supports the desired result when a Developer commits a breach for which its Counterparty can readily prove its damages (e.g., when a Developer fails to pay after recognizing windfall profits).

Reflecting on this point a bit more, I'd be interested to understand how different jurisdictions would handle a situation in which a Developer manages to breach in a more subtle way, before its Counterparty suffers such obvious damages.

For example, suppose that a sophisticated shareholder in Developer A believes that one of A's business units has a meaningful chance (say, 30%) 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 that business unit to a second corporation, B (which does not have a Windfall Clause) in exchange for stock in B, with the stock in B distributed to A's shareholders on a pro rata basis as a dividend. 

Developer A's Counterparty could argue that this sale is a fraudulent transfer and a breach of the Clause. However, to recover damages for that breach, the Counterparty might have the burden of proving (i) the value of the business unit as of the time of the transfer, and (ii) that the value was high enough to implicate the Clause. A court might be reluctant to credit an argument that a single business unit will soon be worth >1% of GDP, limiting the Counterparty's recovery. 

(If enforceable, a liquidated damages provision might help mitigate this concern, but the Counterparty would likely still prefer specific performance.)

Thanks again for taking this on. I'm excited to follow the sequence!

Curated and popular this week
Relevant opportunities