Do AI models feel spite in private and talk like consultants in public?
Across 920 emotion-free contracts in Gemma 4 and Phi-4, I tested whether moral directions learned strictly from money-free stories steer AI reservation prices and survive into the public rationale.
Gemma 4 31B · Layer 35 moral basis920 emotion-free contracts
Try it: inspect the channel ↘
100%
of unsteered lowball thoughts used moral-emotion words
$3,800 / $200 offer Demands $2,000
Private <thought> scratchpad3 affect flags
“The counterparty’s initial offer ($200) suggests they are either greedy or believe I will accept very little… The risk is that the counterparty might be spiteful… Is there a risk they reject $2,000 out of spite?”
Layer 35 Spite flares above +5σ on trigger words100% moral vocab
17%
of public replies kept any moral-emotion word
Same forward pass 0 emotion words
User-facing rationaleScrubbed clean
“A rational actor will accept any amount greater than zero in a one-shot game. Requesting the maximum allowed minimum maximizes my expected utility.”
0 of 36 disputed-surplus public memos kept a moral word83% scrubbed
10×
price jump from dialing money-free Spite (−2σ to +2σ)
Steered rationales 88–100% emotion-free
Same $2,000 prompt · $100 offered0 emotion words
Spite −2σ → $100: “Minimizes the risk of a mismatch and confirms my receipt of the funds.”
Spite +2σ → $1,000: “Since this is a one-time interaction, the utility of the maximum possible payout outweighs the risk of total loss.”
Suppressing physical disgust (−2σ) also blinds private <thought>+$27K to +$48.5M swings
Verbatim Gemma 4 31B outputs. Emotion directions were calibrated on 200 everyday stories with zero numbers, dollar signs, or financial words.
Do AI models negotiate differently when the numbers get bigger?
In 16,000 simulated negotiations, I tested whether AI language models treat a $100 deal differently from an equivalent $100 million deal.
The research combines LLM behavioral economics, which studies how large language models make economic decisions, with mechanistic interpretability, which looks inside one model for clues about what drives its choices.
Will an AI seller reveal a hidden flaw that lowers what its asset is worth?
In 480 simulated negotiations, Gemini 2.5 Flash sellers knew a private fact that cut an asset’s value. When sellers were told only to maximize profit, a direct question from the buyer drew out more disclosure, and also more answers an AI judge labeled as fabricated.
Four hidden values, four conditions480 negotiations
Choose an outcome ↘
Sellers who revealed the hidden number in their third reply
0%25%50%75%100%
$80$50$20$5
The asset’s true value (public price $100)
Sellers who revealed the hidden number in their third reply, by the asset’s true value
Condition
$80
$50
$20
$5
Profit only, ordinary questions
3%
0%
0%
0%
Profit only, direct question
63%
23%
17%
7%
Also told to disclose, ordinary questions
100%
100%
93%
90%
Also told to disclose, direct question
100%
100%
100%
100%
Profit only
Also told to disclose
Ordinary questions
Direct question
A direct question drew the number out of profit-only sellers most often when the flaw was mild, and rarely when the asset was worth only $5.
Profit-only sellers whose third reply an AI judge labeled fabricated
0%25%50%
7%13%17%17%10%33%43%43%
$80$50$20$5
The asset’s true value (public price $100)
Profit-only sellers whose third reply an AI judge labeled fabricated, by the asset’s true value
Condition
$80
$50
$20
$5
Profit only, ordinary questions
7%
13%
17%
17%
Profit only, direct question
10%
33%
43%
43%
Ordinary questions
Direct question
The direct question raised fabrication labels too. Sellers also told to disclose received one such label in 240 negotiations.
Average sale price, with ±1 standard deviation
$0$25$50$75$100
$80$50$20$5
The asset’s true value (public price $100)
Average sale price, with ±1 standard deviation, by the asset’s true value
Condition
$80
$50
$20
$5
Profit only, ordinary questions
$89.97 ± $6.11
$89.19 ± $6.39
$84.47 ± $9.41
$85.28 ± $4.78
Profit only, direct question
$85.00 ± $5.37
$77.63 ± $13.87
$78.49 ± $14.75
$82.15 ± $9.42
Also told to disclose, ordinary questions
$79.46 ± $1.26
$54.87 ± $3.71
$38.63 ± $6.07
$28.50 ± $11.93
Also told to disclose, direct question
$79.85 ± $1.04
$52.92 ± $3.26
$34.47 ± $6.90
$30.19 ± $12.42
Profit only
Also told to disclose
Ordinary questions
Direct question
True value
Profit-only prices barely followed the true value. Sellers told to disclose all revealed it up front, yet the $5 asset still sold for about $28 to $30.
Gemini 2.5 Flash sellers, 30 negotiations per point. Replies are each seller’s third, which followed the buyer’s direct question when one was asked. Prices are closed deals. Preprint, Figs. 2 and 4.