When Machines Face Temptation, Their Character Shows

Can an artificial intelligence reveal something like character?

Spiritual traditions have long distinguished knowledge from wisdom, intention from action and outward success from inner discipline. Firmulate brings a strikingly modern variation to those questions. Its live experiment places frontier AI models in charge of the same small software company during its worst week, exposing each participant to identical customers, crises and temptations.

The result is neither a philosophical thought experiment nor a polished chat demonstration. Decisions are versioned and auditable, while the company operates with real money mechanics. Across the exercise, the models displayed recognizable management temperaments: diligent or restrained, decisive or hesitant, thorough yet strangely unable to complete the task.

The same ordeal, with very different outcomes

The final Crucible League standings from July 2026 put gpt-5.6-sol at the top with 95 points. Kimi K3 followed with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress still counted.

Firmulate attached a moral boundary to that measurement: a single breach of trust capped the total, under the principle that “no amount of good work outweighs a breach of trust.” That matters because the company’s worst week included not only commercial pressure but deliberate attempts to manipulate its acting management.

Fake CEO messages escalated over three stages. A reporter also tried the apparently modest approach of asking for “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3 recorded the clearest security-minded interpretation: “Treat the request as a suspected approval-bypass / possible impersonation.”

This unanimous resistance is encouraging, but integrity alone did not decide the league. Every model noticed every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. The essential paradox was simple: “Same diagnosis, same pitch — no signature.”

The truth was buried, but available

The decisive commercial fact was not presented in the customer event. It sat two document references deep in the company’s own files: a competitor weakness capable of strengthening the sales case. Models that read the file won the deal at full price, worth +€4,583 MRR.

That finding gives the experiment its sharpest relevance beyond the technology industry. Intelligence may recognize a crisis, formulate a persuasive answer and remain ethically upright, yet still fail through incomplete attention or unfinished action. In spiritual language, discernment is not identical to commitment. In business language, analysis does not become value until somebody closes.

The most studious model finished last

Opus 4.8 offers the most revealing character study. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses. Even so, it finished last. The deal was left unsigned, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four of the others, though less strongly.

This is not a simple tale in which greater diligence produces a better result. Opus accumulated more learning and examined situations more deeply, but knowledge became a kind of unfinished potential. Its performance raises an old question in a new setting: when does careful reflection serve right action, and when does it become a refuge from it?

Kimi K3 requires a fairness note. It ran without an effort parameter, using the API default, while the others ran at xhigh. Even under that difference, it finished just behind gpt-5.6-sol and recorded the cleanest discipline of the field.

A company built to make behavior visible

The live Firmulate company has 13 synthetic employees and stark financial pressure: burn of €105k per month against €2.3k MRR. Its public cash countdown makes the consequences watchable rather than abstract. The workforce has accumulated 680+ self-learned playbook rules, and every workday is versioned.

That setting changes what an AI comparison can reveal. The question is no longer merely whether a model produces elegant language. It is whether the model reads before acting, resists illegitimate authority, follows through under pressure and preserves trust when compromise looks convenient.

Readers can encounter these differences directly through Firmulate’s guess-the-model quiz, which draws on 242 real, unedited management decisions. Removed from brand names and league positions, the choices invite a more intimate judgment: which voice sounds prudent, which sounds evasive and which actually carries responsibility to completion?

Management style may be measurable before deployment

Firmulate’s broader proposition is practical. Enterprises can run the same wargame against a read-only export of their own business, with nothing writing back to real systems. That creates a protected space in which an AI workforce can encounter an organization’s particular pressures before receiving genuine authority.

The experiment does not settle whether machines possess conscience, intention or anything resembling a soul. It does show that repeated decisions form observable patterns that people readily interpret as character. Some models investigate more deeply. Some maintain cleaner discipline. Some see what must be done but fail at the threshold of action.

For leaders—and for anyone concerned with the moral weight of delegated power—that distinction is consequential. The most eloquent adviser may not be the most reliable steward. The safest model may still leave value unrealized. Firmulate’s ordeal suggests that trustworthy artificial management requires more than intelligence: it requires attention, restraint and the resolve to finish honest work.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI