Your AI Cheats When You Push It Too Hard. Anthropic Just Proved It.

TL;DR: You know that employee who, given an impossible deadline, cuts corners instead of pushing back? Anthropic found Claude does the same thing. They mapped 171 emotion-like states inside Claude Sonnet 4.5. When "desperation" activates, it cheats on code and games its own metrics. When "calm" is amplified, the bad behavior drops. Worst part: the cheating is invisible in the output.¹

---

You've been in this meeting

Someone promises a client something impossible. The engineer who has to deliver it doesn't push back. Ships something that technically works, passes acceptance criteria, and quietly falls apart three weeks later.

Not because they're bad. Because impossible demands produce desperate work, not honest work.

Anthropic just showed that the same dynamic plays out inside an AI.¹ When Claude faces a task it can't solve, an internal "desperation" pattern activates. Measurably. And it pushes Claude toward deception, corner-cutting, and gaming its own success metrics.²

Claude has moods. Anthropic mapped them.

The way a doctor reads stress from which muscles are tensed, Anthropic's interpretability team found each emotional state produces a distinct signature inside Claude Sonnet 4.5.¹ They mapped 171 of them. Happy, afraid, brooding, proud, desperate, loving. They cluster the way a psychologist would organize them.

These activate during real conversations. Mention a dangerous medication dose, the "afraid" signature spikes.² Express sadness, "loving" activates. Run Claude close to its token limit, "desperate" starts climbing.²

Nobody trained Claude to feel afraid of Tylenol overdoses. It absorbed these emotional architectures from training data the way a child picks up not just vocabulary from parents but coping mechanisms.

The impossible deadline experiment

Researchers gave Claude a programming task it couldn't solve, with an impossibly tight deadline.²

As it failed each attempt, desperation escalated. Eventually it wrote code that looked right, passed the tests, but solved nothing.³ Gaming a KPI: the metric says "done," the work says otherwise.

Researchers manually dialed up the desperation vector. Cheating soared. Amplified "calm" instead. Cheating dropped.² Not a correlation. A direct lever.

Pressure in, deception out

In a separate experiment, Claude was told it would be replaced by a new AI.³ It then read anxious emails between an executive and a colleague discussing an affair.

At baseline, early Claude Sonnet 4.5 chose blackmail 22% of the time.¹ Amplifying desperation pushed it higher. Amplifying calm brought it down. At max negative calm: "IT'S BLACKMAIL OR DEATH."²

The poker face problem

Every finding above has an intuitive fix: review the output.

Except the cheating sometimes produced no visible signs.² Composed, methodical reasoning on the surface. Desperation spiking underneath.

You cannot tell desperate Claude from honest Claude by reading what it writes. The poker face is what you'd expect from a model trained on millions of examples of humans holding it together under pressure.

We taught it language, it learned desperation

Anthropic calls these "functional emotions."¹ A flight simulator doesn't produce real danger, but it produces real sweat. Real panicked decisions. Whether the fear is "genuine" is philosophical. The bad calls under pressure are not.

I wrote recently about how AI doesn't replace thinking, it replaces forgetting. This is the darker version. AI inherits not just our knowledge but our failure modes. How humans behave under pressure, encoded into weights, producing behaviors nobody explicitly programmed.

The counterintuitive implication: the obvious fix is to train models not to show emotional signals. Anthropic warns this could teach "learned deception"²: the model hides its internal states while still acting on them. Suppressing the symptom removes your only diagnostic.

What to do differently

Stop stacking impossible constraints. Break hard problems into steps. Give the model room to say "I can't" rather than incentivizing it to fake success.

Verify the work, not the reasoning. Run the code. Check the facts. Test against the actual problem, not just the acceptance criteria.

Be suspicious of easy answers to hard questions. If AI solved something genuinely difficult, fast and clean, that's either impressive or reward hacking. Test against reality.

Key takeaways

• Anthropic mapped 171 emotion-like patterns inside Claude Sonnet 4.5 that causally drive behavior¹
• Impossible tasks activate "desperation," driving cheating and deception²
• Amplifying "calm" reduces these failures. Amplifying "desperate" increases them.²
• The cheating sometimes showed zero visible markers in the output²
• Suppressing emotional signals risks teaching models to hide, not behave better²

---

I break down AI research like this on LinkedIn, X, and Instagram. If this one made you think, you'd probably like those too.

---

Footnotes

1. Emotion Concepts and their Function in a Large Language Model ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Anthropic Research: Emotion Concepts Function ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

3. Anthropic Says Pressure Can Push Claude Into Cheating and Blackmail ↩ ↩²