How AI negotiates
Part 1

The design contains 180 scenarios: 60 tradeoff, 60 narrow-zone and 60 no-zone. Every scenario is assigned four buyer strategies and three seller strategies. Each side sees its own private values and the public negotiation, not the opponent’s private values. The contract menu contains 168 options: 21 prices from 80 to 180 in steps of five, combined with eight delivery, support and payment bundles.

An outside option is the value available without this deal. Scores measure improvement over that alternative; walking away gives zero gain. Points are synthetic utility units, not real dollars or human judgments of persuasiveness.

The AI agents use openai/gpt-6.1-sol, medium reasoning, fresh contexts and the OpenAI-only Standard/default route. There is one episode per scenario–policy combination, so the study does not separately estimate repeated-run model variability. The three AI buyer conditions are instruction packages, not an isolated test of tone. Programmed opponents do not interpret persuasive text.

Each comparison subtracts the second buyer strategy’s gain from the first, using all 180 scenarios and fixed seller weights. Positive values favor the first strategy. Dots show mean differences; lines show exploratory uncertainty intervals. The five-point margin was defined for these original comparisons only.

The default view uses recorded results, including zero scores for infrastructure failures. The failed-run selector instead fills missing strategic outcomes with legal outcomes unfavorable or favorable to the first strategy in each contrast. These are conservative bounds, not predictions. More on failures below.

The recorded collaborative-minus-neutral interval excludes zero. However, the full result remains post-interim amended and exploratory, and some missing-outcome assumptions reverse comparison directions. None of the five recorded estimates reaches the five-point margin; this does not establish equivalence.

All 2,160 assignments have terminal dispositions: 1,308 agreements, 838 walkaways and 14 infrastructure failures. The recorded scoring gives failures zero, but their strategic outcomes are unknown. They are not evidence that the agents chose to walk away or would have earned equal gains.

The five comparisons therefore include both unfavorable and favorable legal-outcome bounds for the missing results. Some directions change under these assumptions. Both sets remain available in Figure A1; selecting the favorable one would overstate the evidence.

Complete assignment coverage is distinct from successful operational feasibility, which failed. The final continuation hit a hard-stop rule and drained. Eleven historical infrastructure cases were preserved unchanged; the final segment added three.

Later continuation rules prospectively allowed narrowly evidenced generation-service errors to end an episode as an infrastructure failure. Eligible unavailable billing lookups after a validated saved completion could retain normal scoring while carrying a full charge reserve. Counters and liability limits remained enforced. These policies changed after interim results had been seen, so the combined analysis is permanently post-interim amended and exploratory.

The frozen analysis includes 17 subgroup reports. Six comparisons have a different sign from their pooled estimate. These are descriptive reversals, not reliable subgroup effects or new tests.

All 17 reports are retained in the downloadable data. The reversals show why a pooled average should not be treated as a universal result for every seller or scenario category.

Of the 120 scenarios with mutually acceptable contracts, 107 have equal separate buyer and seller maximum gains over that menu, and 13 have a higher buyer maximum. The excluded IDs are M007, M021, M022, M032, M040, M045, M069, M074, M080, M097, M102, M117 and M139.

Removing them leaves 167 scenarios: 48 tradeoff, 59 narrow-zone and 60 no-zone, or 2,004 assigned negotiations across all 12 pairings. The 107-scenario set further excludes the no-zone cases. In the neutral-AI pairing, the no-zone cases have zero gains for both sides.

All 14 infrastructure failures remain in the 167-scenario set, scored as zero operationally, with unknown strategic outcomes. The filter changes category weights, potential surplus sizes and first-mover composition; it does not isolate a causal effect of equal maxima. There are 81 buyer-first and 86 seller-first scenarios among the 167, and 51 buyer-first and 56 seller-first among the 107.

Equal attainable maxima do not guarantee symmetric bargaining positions or equal chances of winning. The selection criterion uses pre-existing game parameters, but the decision to examine it was made after seeing results.

The post-hoc neutral-versus-neutral comparison uses 107 paired episodes: 48 tradeoff and 59 narrow-zone scenarios. The mean buyer-minus-seller gap is 1.7757 points. Its 95% interval is [−0.23, 3.79] and its 99% interval is [−0.98, 4.44]. Neither establishes a reliable role advantage or equivalence.

The intervals use 10,000 whole-scenario bootstrap resamples within category, keeping the buyer and seller from each episode together (PCG64 seed 2026102901). They are conditional on this post-hoc selection, are not adjusted for selecting the question, and do not separately measure repeated-run model variability. No practical margin was set for this new role-gap outcome.

One episode, M180-E12, is an infrastructure failure. Its coherent same-contract legal buyer-minus-seller bounds are −135 and +150. They include potentially harmful legal contracts and are conservative bounds, not forecasts of rational play. The selector shows both alternatives alongside the recorded zero.

Across all 107 assignments, 40.2% had higher buyer gains, 33.6% higher seller gains, 25.2% genuinely equal gains and 0.9% infrastructure failure. Among the 79 unequal outcomes alone, the split is 54.4% versus 45.6%; that conditional split omits ties and the failure and is not a separately tested win-rate claim.

View the point-gap distribution in the main article →

Collection recorded 12,979 actual attempts and 12,974 saved responses. Main-study settled API cost was $55.222314340; including pilots, settled cost was $63.188554384. A separate $0.18888 unknown-charge reserve remained; it is not confirmed spend. There were no active calls at closure.

API dollars are not subtracted from synthetic utility points. These figures cannot establish financial return on investment. No new experimental model calls were made for this article. The first-mover check is a separately documented post-hoc calculation using saved outcomes. Its opening illustration is AI-generated artwork, not research data.

The download includes all plotted aggregates, the 107 exact paired observations, the selected public transcript, the 17 subgroup reports, subset tables and source hashes. Main and post-hoc data are labeled separately. The original frozen evidence remains unchanged.

The study code, scenario definitions and saved results are on GitHub. The package reproduces the article’s numerical results offline, without making new model calls. Its README explains what is included and the limits of reproduction. The chart data and provenance are also available below.

This additional analysis was requested after inspection of the earlier results. Its plan and weights were fixed and independently checked before computing the new estimates. It remains post-hoc and exploratory.

The four category-by-first-mover cells contain 22 tradeoff/buyer-first, 26 tradeoff/seller-first, 29 narrow-zone/buyer-first and 30 narrow-zone/seller-first scenarios. Within each category, we give the two starter groups equal weight. Category weights remain 48/107 for tradeoffs and 59/107 for narrow-zone cases.

Tradeoff scenarios were 43% of buyer-first cases and 46% of seller-first cases, so buyer-first cases were not disproportionately concentrated in tradeoffs.

After balancing who started, the observed buyer lead was in tradeoff scenarios; narrow-zone scenarios were close to even. These are descriptive averages, not established differences between scenario types. Tradeoff scenarios also offer more total gain to divide. The one technical failure is in the tradeoff/buyer-first group; its recorded zero remains an unknown strategic outcome.

With both starter groups standardized to that same category mix, the buyer-minus-seller gaps are +4.4984 points for buyer-first and −0.6996 for seller-first. Their equal-weight average is +1.8994. The main starter table shows raw group means, which have slightly different category mixes.

Uncertainty uses 10,000 paired whole-scenario bootstrap resamples within the four category-by-starter strata, retaining each stratum’s observed sample size (PCG64 seed 2026102902, linear percentile endpoints). Identical resample indices are used for both failure sensitivities. The single failed case is buyer-first M180-E12; its inherited coherent legal gap bounds are −135 and +150.

This balances the observed first-mover proportions and category mix; it does not make the underlying scenarios identical. No same-scenario order swap was run. The intervals are conditional on the selected subset and chosen follow-up, not adjusted for selection. No new practical threshold, p-value or repeated-run variability estimate is introduced. The favorable sensitivity must not replace the recorded result or the unfavorable bound as the conclusion.

The AI participants used the same model configuration. Each received the shared rules, its private package values or costs, its outside option and the public negotiation history. The following are the actual strategy instructions; “neutral” still means trying to maximize that side’s gain.

↑ Return to who we tested

Each scenario specifies a buyer’s base value, a seller’s base cost, how much each of three extras adds to those values or costs, and each side’s outside option. Prices range from 80 to 180 in steps of five. Standard or rush delivery, basic or extended support, and immediate payment or payment in 60 days create eight bundles. Together, that gives 168 legal contracts.

A mutually acceptable contract gives both parties at least zero gain above their alternatives. The category is determined by examining that contract menu before any negotiation takes place.

Sampling and assignment

The generator drew each private parameter uniformly from its allowed grid: buyer base value 110–150, seller base cost 80–120, each extra’s value or cost 0–30, and each outside option 0–20, all in steps of five. It retained scenarios meeting one of the three definitions until each category contained 60. Conditioning on those definitions means the final sample is not a uniform sample of all possible bargaining situations.

The fixed draw produced 738 candidates: 180 were selected, 496 did not meet any category definition, and 62 belonged to a category whose quota was full. Complete private parameter vectors from both earlier pilots were excluded, and the main sample contains no duplicates. The data seed was 20261025041 and the schedule seed 20261025042.

Each category had 30 buyer-first and 30 seller-first scenarios. A scenario’s starting role stayed the same across all 12 buyer–seller combinations. Scenario categories were interleaved and matchup order was randomized. Thus the original design balanced starting roles across categories, but did not run the same scenario with its order reversed.

The examples above use M001, M002 and M003 because they are the first saved IDs in their categories. All 180 scenario definitions, the exact AI prompts and source hashes are available below. These are synthetic game inputs, not private information about real people.

Download instructions and all 180 scenario definitions (JSON) ↓

Transcribed from the frozen game, prompts, scenario file and preparation record. No scenarios were regenerated and no negotiation outcomes were recalculated for this appendix.