← Evidence briefs

Octopus and bee sentience compared with the case for AI

What this gives us: a working model of how to grade evidence with fixed criteria and confidence levels, set before looking. That's the pattern our own tests should copy.

2026-10-03

Checked by our internal skeptic review; corrections welcome.

Brackets show how each source was read.

Both full-text anchors, the 2021 report and Birch 2024, are led by Jonathan Birch, so this brief leans heavily on one research group's framework. Gibbons, both Solvi papers and Baracchi were read as abstracts only.

1. The markers

Birch and four co-authors wrote the 2021 report via LSE Consulting; Defra, the UK environment department, chose the animals [full text]. The report used eight criteria:

1. nociceptors (damage-detecting nerves);
2. integrative brain regions;
3. connections between the two;
4. responses changed by painkillers;
5. trade-offs between threat and reward;
6. flexible self-protection, such as wound-tending;
7. associative learning beyond habituation;
8. an injured animal valuing a painkiller.

Grades ran from "very high confidence" to "no confidence." The report stresses that low confidence does not mean sentience is unlikely or disproven. High or very high confidence on seven or more criteria meant very strong evidence; on five, strong; on three, substantial.

2. Octopus

• The report: "There is very strong evidence of sentience in octopods." Confidence was high or very high on seven criteria and only medium on trade-offs [full text].
• Crook (2021, iScience) is one small study. Octopuses avoided a chamber linked to an acid injection, preferred one linked to anaesthetic relief, and stopped grooming the wound after anaesthetic. The study used 8 injected animals and 7 controls, in one species. Crook concludes: "Thus, octopuses are likely to experience the affective component of pain." She notes the anaesthetic's effect on place preference was variable, possibly from incomplete or off-target injection [full text].
• What is debated: Crook notes ongoing debate about how behaviour links to inner feeling [full text]. We did not search for replications.

3. Bees and other insects

• Gibbons et al. 2022 (PNAS): Bumblebees accepted harmful heat for sweeter sugar water. The authors write: "This ability is consistent with a capacity for pain experiences in insects." [abstract] Birch notes that three authors, including Crump (a 2021 report co-author), are associated with his project; he was not involved [full text].
• Solvi, Baciadonna & Chittka 2016 (Science): a positive emotion-like state. After an unexpected sugar reward, bees responded more positively to ambiguous cues, and a dopamine blocker removed the effect [abstract]. A separate study, Solvi et al. 2020, looked at cross-modal recognition [abstract].
• Galpayage Dona et al. (2022): bees rolled balls without reward. The authors suggest the behaviour may reflect positive affective states. They also write: "We cannot exclude that object play in bumble bees occurs following a lack of stimulation that would otherwise be provided by the natural environment." [abstract and excerpts of the authors' PDF]
• What is debated: In a 2017 review (not an experiment), written before the 2022 studies above, Baracchi, Lihoreau and Giurfa write: "We maintain that caution is warranted before attributing emotion-like states to honey bees and bumble bees as some experimental caveats may undermine definitive conclusions." [abstract]
• Birch's view is that bees are sentience candidates. He calls no single result conclusive but judges that bees and fruit flies already qualify. His definition reads: "A system S is a sentience candidate if there is an evidence base that (a) implies a realistic possibility of sentience in S that it would be irresponsible to ignore when making policy decisions that will affect S, and (b) is rich enough to allow the identification of welfare risks and the design and assessment of precautions." [full text]

4. What this means for AI

Markers that can be gamed have to be discounted. That warning, and the design rules that make faking harder, are on Hardest to Fake. This page does not repeat them.

[OUR REASONING] Anatomical markers like nociceptors do not transfer to AI. Missing them does not by itself rule AI out, though theories tied to biology would count it against. Intelligence, learning or cognition would not count alone.