My apologies @Vasco Grilo🔸 I was lazy there and didn't read the whole article properly. Thanks for pointing out the important areas.

No worries. I could have highlighted them more too. So thanks for asking. I linked to the relevant section in this comment, but I have now linked to it at the top of the post too. "The article has a section discussing promising ways of assessing sentience (here is some brief context [my comment just above])".

Nitpick. I get email notifications of comments on my posts, and replies to my comments even if I am not tagged. I think this is the default setting on the EA Forum. If I am tagged in those cases, I get 2 email notifications, although this does not bother me.

Working memory and operant conditioning don't carry any weight with me personally, I struggle to connect those with a meaningfully increased chance of pain experience.

I would also give them the least weight among the 4 indicators, but I am not confident they should have negligible weight. I think negligible weight makes more sense for classical/respondant conditioning, as SPUDs are capable of this.

In addition, I believe some types of working memory and operant conditioning should have more weight than others. Here is some brief context from the post.

More stringently, it could be that not all operant learning requires sentience [it demonstrably does not require consciousness of the rewards and punishments], but that only some sub-types do. Thus, it could be that Ginsburg and Jablonka's (37) hypothesis about second-order conditioning [context] is correct: again something not yet known, but amenable to empirical test. If their hypothesis is supported [it may not be], then again there are already some cases demonstrating conditioned responses to secondary reinforcers by fish.

These higher order processes would be most convincing for me than the memory stuff, especially deeper levels of self recognition which I feel like could maybe be tested? 

Perhaps a silly question, but I thought the mirror/parasite mirror test was related to "self recognition", but you seem to mention that as something different.

Both the mirror test and “mirror test with biting parasite” test self-recognition, but the latter offers stronger evidence for sentience. I asked Gemini 3.5 Flash to contrast the 2 tests with respect to this given the article's description of the “mirror test with biting parasite”. The answer is below. I think it is good, except that it is overconfident (as it is often the case with LLMs). Words like "proves" and "demonstrates" should be replaced with "offers stronger evidence (relative to the standard mirror test) for".

The standard mirror test is widely used to check for self-awareness, but it doesn't tell us if an animal can actually feel things like pleasure or pain.

Mason and Lavery (2022) proposed a modified version to bridge this gap. Here is how the two tasks contrast and why the new one offers stronger evidence for sentience.

The Two Tasks Contrasted

• The Standard Mirror Test (Affectively Neutral): An animal is marked with a spot (like a dot of odorless, painless paint) in a place it can only see using a mirror (like its forehead). If the animal looks in the mirror and then grooms, touches, or scrapes its own body where the mark is, it passes. This proves the animal understands that the reflection is "me" rather than another animal, showing a baseline of physical and visual self-awareness. However, because the mark doesn't feel like anything, the task is affectively neutral—it doesn't involve emotions or feelings.
• The "Mirror Test with a Biting Parasite" (Affectively Charged): This version introduces feelings. Instead of a neutral dot, the animal is trained to associate different visual marks with physical comfort or discomfort. For example:
  1. The animal is exposed to a black triangle mark that is physically uncomfortable or painful (simulating a biting parasite).
  2. It is also exposed to a white circle mark that causes no discomfort at all.
  3. Finally, the researchers place these marks on the animal's body without any actual pain or discomfort attached, in a spot the animal can see directly or via a mirror. The test checks if the animal reacts with greater urgency, anxiety, or scraping behaviors toward the "nasty" black triangle than the "benign" white circle, purely based on what it looks like.

Why the Parasite Test Provides Stronger Evidence for Sentience

To understand why the parasite version is stronger, it helps to separate two concepts: awareness and sentience.

Passing the standard mirror test proves awareness—specifically, the cognitive ability to map visual reflections to one's own physical body. But critics argue that a highly complex, non-sentient robot or AI could theoretically pass this test using pure geometry and visual processing, without experiencing any inner feelings.

The biting parasite test targets phenomenal consciousness (P-consciousness) and affective states—the actual capacity to subjectively feel pain, fear, or distress.

By showing that an animal doesn't just recognize a mark, but specifically anticipates and reacts to the remembered emotional and painful quality of that mark, the test demonstrates that the animal experiences the world through a lens of "good" and "bad" feelings. It links visual self-recognition directly to an inner emotional state (the desire to avoid pain), providing much sturdier proof that the animal is truly sentient and capable of suffering.

Hi Nick.

What kind of research do you think could decrease uncertainty about their sentience?

The kind of research described in this section of the article. Tasks involving working memory, aspects of operant conditioning, self-report of sensation, or higher order consciousness (like the “mirror test with biting parasite”, which offers stronger evidence for sentience than the standard mirror test). Below is some context from the article about why these tasks are promising. I bolded the 4 types of tasks.

here we suggest four types of candidate measure that seem particularly promising as indicators of sentience. One concerns working memory, and another, forms of operant conditioning: two types of ability that S.P.U.D. subjects seem not to convincingly have. A third type of candidate measure concerns self report. As we saw in the section Red Herrings: Responses That Do Not Require Sentience, unaware humans, and also monkeys with blindsight, report sensing nothing (thence having no P-consciousness), even when presented with stimuli that elicit other kinds of response – with the human subjects consistently describing themselves as merely guessing. Finally, higher order cognitive processes such as episodic memory and self-recognition, can yield insights into animal awareness. These have not been studied in S.P.U.D. subjects, and this would likely be impossible, but are arguably reliant on perceptual consciousness. We review each of these below, and also outline how they could be modified to now address questions about sentience: the ethically relevant dimension of P-consciousness at the heart of the fish pain debate. We also suggest how their validity as indicators of sentience could be assessed.

In addition, tasks involving endogenous/voluntary attention, like Posner’s spatial cueing task, as discussed in Nieder (2022).

I'm assuming you mean behavioural research?

All of the above tasks look into behaviour. However, passing them would be stronger evidence for sentience than passing tasks that have been passed by SPUDs (spines disconnected from brains, humans in unaware states, decerebrate mammals and birds, or organisms lacking a nervous system like plants and protozoa).

I think this kind of research is helpful to reduce uncertainty but doesn't necessarily address fundamental questions about how brains work.

Note that research looking for the neural correlates of consciosness (NCCs) in animals relies on studying the behaviour of humans. In particular, to understand which states of the human brain are responsible for consciousness. Nieder (2022) says there is no evidence for the neural correlates of consciousness in animals besides mammals and birds.

I'm just saying that if you're not a longtermist, you don't face as much uncertainty about how to achieve good outcomes

Makes sense.

Not sure what you're asking exactly

You seem to suggest that longterm effects may not be relevant for global health interventions while being relevant for AI safety interventions (or others which are typically referred to as longtermist interventions). I meant to question this. If I thought longterm effects were relevant for AI safety interventions (I do not), I would think they would be relevant for global health interventions too.

I think the other lists have value, but are more more observational heuristics that might explain many micro issues, but not really help you make sense of the world and how things work overall.

I agree.

Hi Elias.

We face a lot of uncertainty about the sign of our impact.

Therefore, we should be very vigilant about our epistemics to make sure that we are not having a negative impact in expectation.

But trying hard deeply distorts our epistemics - it makes us more prone to motivated reasoning about what we’re doing, and leaves us with less slack to reflect on it.

Great point.

Crucially, this argument applies much more strongly to people working in “longtermist areas” - which other critiques of trying hard generally don’t do. For example, global health EAs whose terminal value is short-term welfare also face uncertainty about the impact of their actions - but much less (especially about the sign) than people trying to improve the long-term future.

If interventions decreasing the risk of large catastrophes mostly affected the longterm future, why would the same not apply to global health interventions?

Interventions with negligible longterm effects could still decrease welfare due to effects on soil invertebrates?

Hi Matt. Nice initiative.

Do people come away understanding these 200 concepts well?

Here are some related lists.

Hi Peter.

TL;DR: We ran a Delphi study with 272 international AI experts to prioritize 24 AI risk domains from the MIT AI Risk Domain Taxonomy.

I wonder how much the probabilities would change if they were about all sources of risk instead of just AI. If they would not change much, could it be that AI has not increased the risks, but simply became associated with all risks because it is in the process of being integrated into practically everything that is relevant for assessing the risks you covered?

Hello.

Based on everything you've looked into, what odds do you put on fish sentience?

I speculate carp have around 50 % (= (0.70 + 0.40)/2) chance of being sentient, but I have little reason to expect my intuitions are calibrated. I feel like anything from 0.1 % to 90 % is reasonable. In any case, I can see the welfare of fish being very close to 0 even if they are sentient. So I would rather prioritise decreasing uncertainty about their sentience and intensity of their experiences over investing in interventions helping fish.

Here is some additional context you may be interested in about the likelihood of fish being sentient. Bob Fischer's book about comparing welfare across species presents an estimate of 70 % for the sentience of carp. Among the people surveyed in Zipple et al. (2024), as illustrated in Figure 1 below, around 40 % guessed "most" or "all or nearly all" fish are conscious. Importantly, "we defined consciousness for respondents, using a ‘self-consciousness’ or ‘self-awareness’ definition (an animal being aware of its own existence [33–35])".

I would add, I'm not convinced that all "unconscious" states in humans are equal. I've heard reports of people who "wake up" during surgery, but were still unable to move.

Anesthesia awareness is a real (though rare) phenomenon. However, note that people who are awake would no longer qualify as unconscious under the definition of the article.

Thanks for the good points. Here is the same graph for the United States (US). The income before tax of the 1 % of people with the most income was supposedly 20.4 % in 1913, and 20.7 % in 2024. I am sharing data for the US because it covers a long period (111 years), and "has been the world's largest economy since about 1900" until 2015.

I remain open to bets against short timelines for transformative AI (TAI), or what they supposedly imply, up to 10 k$. Do you see any that we could make?

Spencer Greenberg has a related essay. 50 “Laws” of Everything.