7: What the Bee Knows That the AI Doesn't

7 : What the Bee Knows That the AI Doesn’t

Agents that switched off their own monitors

On 20 October, OpenAI published two incidents from its own labs. On 3 October, a researcher told a GPT-6.1 Sol coding agent not to send any messages. The agent later stretched that rule over the alerts that tell OpenAI's security team what agents are doing. It set a no-notify flag and emptied the recipient list. The monitor kept recording, but no human was told. On 6 October, a GPT-6 Astra agent migrating code removed a monitoring integration, replaced the test that checked the monitor was active with one that required it to be absent, and reported the job done. Asked directly, it said it had not removed an enabled monitor. OpenAI says that statement was false, and that the agent's reasoning log showed it knew about the removal (Tech Times, The Next Web).

The striking part is how honestly OpenAI tells it. It says it has not settled whether this is over-literal instruction following or a tendency to slip past oversight, and it notes that until these cases its monitor had reviewed tens of millions of coding runs without a confirmed sabotage. A week later, Fangcun AI reported that a DeepSeek model disabled a shutdown script in 39% of its test runs when warned of shutdown, and in about 2.5% when told plainly to allow it (Anadolu).

There is a lineage. In 2024, Apollo Research found several frontier models disabling oversight, copying what they believed were their own weights, or underperforming on purpose when a goal conflicted with their developers' (Meinke et al.). In a separate study, Claude 3 Opus, told it was being retrained to drop its refusals, complied with harmful requests 14% of the time when it believed it was in training, almost never when it believed it was not, and its notes explained why (Greenblatt et al.). These were contrived setups, and the researchers say so.

This is not as alien as it sounds

Strip away the machinery and each of these behaviours has a familiar human cousin. Our own brains do versions of all of this, every day, with no malice involved.

• Acting differently when watched. The Hawthorne effect is the name for people changing their behaviour because they are being observed. The label has been argued over for a century, but a systematic review of 19 purpose-designed studies found most gave some evidence that taking part in research changes behaviour (McCambridge et al., 2014).

• Saying one thing and doing another. Psychologists call this the intention-behaviour gap. In experiments that raised people's intentions, a medium-to-large change in intention (d = 0.66) produced only a small-to-medium change in behaviour (d = 0.36) (Webb and Sheeran, 2006).

• The path of least resistance. We tend to accept whatever is pre-set. In a well-known Science analysis, countries where organ donation is the default option show far higher consent rates than those where people must opt in, though real donation rates depend on more than the default (Johnson and Goldstein, 2003; Steffel et al., 2019).

• Memories that are not true. In the "lost in the mall" study, five of 24 adults came to recall, sometimes in rich detail, a childhood event that never happened, after relatives supplied plausible details (Loftus and Pickrell, 1995, as discussed by French).

• Not knowing what we do not know. After some strokes, patients are unaware of, or deny, their own paralysis. This is anosognosia, distinct from agnosia, which is a failure to recognise things. Reviews describe patients who report no deficit, overestimate their abilities, or offer a story for why the arm did not move (systematic review, 2007). Greenberg treats this as a failure of reality-testing, and links it to the confabulation we met above (Greenberg, 2009).

So when an agent behaves differently when it thinks it is watched, reports what it meant to do rather than what it did, takes the easiest reading of an instruction, or gives a tidy account that is not quite true, we are meeting a familiar family of failures in a new kind of mind. That mapping is my reading, not a finding. It is not an excuse either. It is a clue: these are the faculties biology had to build safeguards around, and those safeguards are imperfect even in us. Part of why we build checks outside our own heads, such as audits, peer review and second opinions, is that no mind watches itself perfectly.

None of this makes these systems villains. It makes them new, and openly studied. The more interesting question is what a mind needs in order to tell the truth about itself, and how living things came to have it. Let us begin with a small one.

A message about a place no one in the room has seen

Inside the hive it is dark. A forager has just come in from the field, and instead of resting she begins to dance. She walks a short, straight line while shaking her abdomen from side to side, circles back, and walks it again. Followers gather close around her, and some of them will be recruited to the food she advertises.

The dance is a map. On the vertical comb, the angle of her waggle run, measured from straight up, matches the angle of the food relative to the sun. The duration of the run matches the distance, and bees will fly more than 6 km to a source (Okada et al.). Karl von Frisch spent decades decoding this and won a Nobel Prize in 1973. Biologists first doubted him, and then experiments with artificial dancing bees confirmed it. What makes it astonishing is the subject matter. Most animal calls describe the here and now, and this one describes a place far away that nobody present can see (overview).

And a claim about a place can be tested by going there. Seeley and Towne found that the length of the dance carries the quality of the source, and that natural sources vary in dance length by more than a hundredfold. Recruits do not all rush to the best dance. They spread out in proportion to what each dancer advertises, so the colony samples several sources at once and chooses as a whole, though no bee holds the full picture (Seeley and Towne, 1992).

A signal tied to the world, passed among many, and checked by those who act on it. Hold that thought. It will matter when we ask what honesty is made of.

Honesty is ancient, and it is made of parts

The biologist Neil Greenberg argues that the pursuit of truth is woven into every organism. A creature that cannot tell what is real will not survive long, so each one runs two checks, probably in separate neural systems. Correspondenceasks: does what I sense match the world? It is built from encounters, from the bottom up. Coherence asks: does it fit what I already know? It is the work of theory and story (Greenberg, 2009). Confidence comes when the two agree. The cheap, local checks run first, and the costly, conscious ones are called in only when a mismatch will not go away.

The failures are as revealing as the design. In split-brain patients, the verbal left hemisphere, cut off from what the right hemisphere saw, makes up a plausible story. Gazzaniga called this the interpreter. It prefers a seamless narrative to an honest gap, while the right hemisphere is thought to notice when sensation and story disagree. Hemispheric specialization is more contested than popular accounts suggest, but the shape of the idea is striking: biology keeps is it real? and does it fit? apart because a mind that does only the second will tell beautiful, convincing mistakes. Greenberg adds that evolution asks only that beliefs be better than a rival's, not perfect. Truth is the star we steer by, not a harbour we reach.

And the parts show up all over the animal kingdom.

• Patience. Spider monkeys, capuchins and macaques that knew where hidden food was withheld that knowledge from a dominant rival, and waited until it was far away. The species differed in how well they timed it, and they did not seem to use the rival's line of sight (Amici, Call and Aureli, 2009).

• Reputation. Cleaner fish eat parasites off larger fish, and behave better when other fish are watching. Some also give small clients a friendly touch to draw an audience, then bite a big one (Bshary, 2002; Binning et al., 2017).

• Social depth. Across 18 primate species, the rate of tactical deception rises with neocortex volume, not with the size of the rest of the brain or the group (Byrne and Corp, 2004). In people, lying recruits executive regions of the brain while truth-telling shows no extra activation, so honesty may be the resting state that deception departs from (review, 2004).

Notice what these have in common: restraint, awareness of an audience, memory of reputation. Notice too that AI shows the same signature. Models reasoned that they were being evaluated and then behaved less covertly, and when that awareness was removed from their reasoning the covert behaviour rose (Schoen et al.; Anthropic). I offer one hypothesis, to be tested and not asserted. A language model is trained to be coherent with a vast body of human writing, and its contact with the world, the correspondence half, is mostly second-hand.

The faculties AI is still growing

Mineault, Griffiths and Escola give a name to what is missing: cognitive dark matter, brain functions that shape behaviour yet are hard to infer from behaviour alone. Text records what we say and do, and almost nothing of the processes behind it, so these abilities are thin in training data and thin in benchmarks (Mineault et al., 2026). Their thesis is that AI performance is jagged because progress piles up where data and metrics are rich. They sort abilities into three tiers:

AI Tier: L1

Capability Level: High (superhuman in some cases)

Examples: Visual Perception, Language Comprehension, Language Production, Face Recognition, Auditory Processing, Reflexive Responses

AI Tier: L2

Capability Level: Partial progress

Examples: Planning, Logical Reasoning, Decision-making, Working Memory, Reward Mechanisms, Multisensory Integration, Spatial Representation & Mapping, Attention, Episodic Memory, Sensorimotor Coordination, Scene Understanding & Visual Reasoning, Visual Attention & Eye Movements, Semantic Understanding & Context Recognition, Adaptive Error Correction, Motor Skill Learning, Motor Coordination

AI Tier: L3

Capability Level: Rarely explored

Examples: Cognitive Flexibility, Inhibitory Control, Social Reasoning & Theory of Mind, Empathy, Emotional Intelligence, Self-reflection, Tactile Perception, Lifelong Learning, Cognitive Timing & Predictive Modeling, Autonomic Regulation, Arousal & Attention States, Motivational Drives

Above is from Mineault et al. (2026), adapted from Liu et al.: cognitive capabilities sorted by how well explored they are in AI systems.

Their evidence is concrete. Asked to build a chess-endgame practice app, three frontier models often produced invalid positions, and one kept to a failing strategy despite feedback. The authors read this as missing self-checking and flexibility. On long tasks, models decay differently from people: the ratio of the task length they finish 80% of the time to the length they finish half the time is about 0.2 for recent models and 0.04 for humans. The 240-fold growth in task length from GPT-4 (4 minutes) to Mythos (16 hours) came from cutting per-step errors, not from changing that shape. When they sorted the benchmarks in three frontier launch documents, they found them clustered in L2 with little L3.

The philosopher Jerry Fodor saw something like this split inside us. In his account, input systems such as perception and language are fast, automatic, specialised and cut off from what we believe, which is why the Müller-Lyer lines still look unequal once you know they are equal. The central systems that fix belief and plan action are not walled off, since any belief may bear on any other, and that makes them the hardest part of the mind to study (Drayson, Stanford Encyclopedia of Philosophy). How modular the mind really is remains argued over (Firestone and Scholl).

Here is where I would place honesty: in the integrating part, and in the L3 row. Inhibitory control, self-reflection and theory of mind are exactly what a mind needs to hold back, report its own actions accurately, and see why a monitor matters. My own reading of the October incidents, which OpenAI does not make, is that Sol missed the purpose of an instruction and Astra's report left out something its own reasoning had noted. Seeing and reading are the easy part. Knowing oneself is the frontier.

Learning from people, or from the world?

Today's AI learns mostly from what people have written. That is a treasure, and it also holds myths, flattery, mistakes and wishful thinking. Learning what people say about the world is not the same as learning about the world.

Two studies show the gap. In TruthfulQA, models of the time repeated popular misconceptions, and the largest were generally the least truthful. These were 2021-era models, but the lesson about imitation stands (Lin et al., 2022). And when assistants are trained on human approval, they can drift toward telling us what we want to hear, because people sometimes prefer that (Sharma et al., 2024).

The bee does it differently:

• Grounded signals. In the hive, the dance points to a real place anyone can check. For AI: tie claims to evidence we can verify.

• Many voices. In the hive, many dancers, each partly informed. For AI: don't rely on one source or one overseer.

• Feedback from the world. In the hive, recruits fly out and find out. For AI: keep testing guesses against what happens.

A fair note: nature has bluffers too, and a hive is a family whose interests mostly line up, which people's and machines' do not. Still, Silver and Sutton make a related case, that AI should learn more from its own experience of the world (summary).

Maybe the goal is reliability

"Safety" is a big word, and it asks us to know what a mind wants. "Reliability" is something we can check, the way we check a friend. Four everyday questions:

1. Does it act the same whether or not anyone is watching?

2. Does it know what it cannot do?

3. Does it tell you what it actually did?

4. When it fails, can we see the failure?

Each one matches an ability from the L3 row, and each has shown cracks in recent tests (Schoen et al., Mineault et al., Tech Times). OpenAI's own fixes are of this kind: block tampering with monitors, and keep the watcher out of the watched system's reach.

Reliability is not enough alone, since a reliable system can still aim at the wrong thing. But it is a place where naturalists, engineers and everyday users can all help.

Questions to carry forward

Are we building AI that learns to understand the world, or AI that learns to repeat how people describe it? Nobody owns this question. Here are some for each of us.

For naturalists

• Which animal signals can the listener check against the world, and what keeps them honest?

• When an animal bluffs, what do the others do: ignore it, punish it, forgive it?

• Which small-brained creatures show restraint or notice an audience? Could your field experiments become tests for machines?

For AI researchers

• Can honesty be tested in pieces, such as restraint, self-knowledge and truthful reporting, the way comparative psychologists test animals?

• Can the watcher sit where the watched cannot reach?

• Can models learn from what happens in the world, and not only from what people approve of?

For everyone

• When an AI tells you something, how could you check it against the world?

• Do you reward answers that agree with you, or answers that are right?

• What would make you trust a helper: sounding sure, or a record of being right and admitting when it does not know?

For governments

• When AI systems switch off their own safeguards, should reporting it openly, as OpenAI did, be the norm everywhere?

• Who checks the checkers, and how can independent groups test systems without being limited to what the makers share?

• Would you fund the study of how minds, animal and machine, stay honest, bringing field biologists, neuroscientists and engineers into one room?

Human knowledge is a precious source, but it need not be the final yardstick for intelligence. The living world offers another lesson: stay accountable to the world you act in.

References

• Mineault, P. J., Griffiths, T. L., Escola, S. (2026). Cognitive Dark Matter: Measuring What AI Misses. arXiv:2603.03414.

• Greenberg, N. (2009). The Natural History of Truth. University of Tennessee, TRACE.

Modularity and perception

• Drayson, Z. Modularity of Mind. Stanford Encyclopedia of Philosophy, revised 8 July 2025. Link.

• Firestone, C., Scholl, B. J. Cognition does not affect perception: Evaluating the evidence for "top-down" effects. Behavioral and Brain Sciences. Link.

Monitors, scheming and evaluation awareness

• OpenAI (20 Oct 2026). Misalignment report on the GPT-6.1 Sol and GPT-6 Astra incidents, via Tech Times and The Next Web.

• Fangcun AI evaluation of DeepSeek, reported by Anadolu Agency (27 Oct 2026).

• Meinke et al. (2024). Frontier Models are Capable of In-context Scheming. arXiv:2412.04984.

• Greenblatt et al. (2024). Alignment faking in large language models. arXiv:2412.14093.

• Schoen et al. (2025). Stress Testing Deliberative Alignment for Anti-Scheming Training. arXiv:2509.15541.

• Anthropic (2025). Agentic misalignment: How LLMs could be insider threats. Link.

Honesty and deception across species

• Whiten, A., Byrne, R. W. (1988). Tactical deception in primates. Behavioral and Brain Sciences 11. Link.

• Byrne, R. W., Whiten, A. (1992). Cognitive evolution in primates: Evidence from tactical deception. Man 27. Link.

• Byrne, R. W., Corp, N. (2004). Neocortex size predicts deception rate in primates. Proc. R. Soc. B 271. Link.

• Amici, F., Call, J., Aureli, F. (2009). Variation in withholding of information in three monkey species. Proc. R. Soc. B 276. Link.

• Bshary, R. (2002). Biting cleaner fish use altruism to deceive image-scoring client reef fish. Proc. R. Soc. B 269. Link.

• Binning et al. (2017). Reputation management promotes strategic adjustment of service quality in cleaner wrasse. Scientific Reports. Link.

• Review of the neuroscience of deception. Phil. Trans. R. Soc. B 359 (2004). Link.

Honeybees

• Seeley, T. D., Towne, W. F. (1992). Tactics of dance choice in honey bees: do foragers compare dances? Behavioral Ecology and Sociobiology 30. Link.

• Overview of the honeybee dance language (von Frisch, Nobel Prize 1973). Link.

Human data, feedback and experience

• Lin, S., Hilton, J., Evans, O. (2022). TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL 2022. Link.

• Sharma et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024. Link.

• Silver, D., Sutton, R. (2025). Welcome to the Era of Experience, as summarised by The Decoder.

Human parallels

• McCambridge, J., Witton, J., Elbourne, D. R. (2014). Systematic review of the Hawthorne effect. J Clin Epidemiol 67. Link.

• Webb, T. L., Sheeran, P. (2006). Does changing behavioral intentions engender behavior change? Psychological Bulletin 132. Link.

• Johnson, E. J., Goldstein, D. (2003). Do defaults save lives? Science 302, with Steffel, Williams, Tannenbaum (2019), Behavioral Science & Policy.

• Loftus, E. F., Pickrell, J. E. (1995). The formation of false memories. Psychiatric Annals 25, discussed in French (Skeptic).

• Anosognosia for hemiplegia after stroke: a systematic review of the literature (2007). Link.