Stated to be Human, Revealed to be Algorithmic: The Trust Paradox Inside LLMs

Trust is a convenient word. Too convenient, really.

In business meetings, people say they “trust the analyst,” “trust the model,” “trust the expert,” or “trust the dashboard,” as if trust were a stable property sitting neatly inside the decision-maker. Then the actual decision arrives, with a deadline, a performance table, a projected loss, and someone quietly asks the AI assistant which source to follow.

That is where the polite trust statement begins to wobble.

The paper behind today’s article, Language Models Exhibit Inconsistent Biases Towards Algorithmic Agents and Human Experts, studies exactly this wobble in large language models.¹ The authors ask a simple but uncomfortable question: when an LLM must weigh a human expert against an algorithmic agent, does it behave consistently?

The answer is: no, not reliably.

When prompted directly to rate trust, the tested mid-2024 LLMs tended to say they trusted human experts more than algorithms. That looks like familiar human-style algorithm aversion. But when the same broad question was converted into a performance-based delegation task — the model sees examples of a human and an algorithm making predictions, then must place a bet on who will be right next time — many LLMs disproportionately chose the algorithm, even when the human had clearly performed better.

So the headline is not “LLMs hate algorithms.” Nor is it “LLMs worship algorithms.” That would be too tidy, and therefore suspicious.

The more useful lesson is sharper: an LLM’s explicit statement about whom it trusts may not predict how it delegates work when evidence and task framing change. In an enterprise workflow, that difference is not philosophical decoration. It is the difference between a chatbot that sounds balanced and an agent that routes a decision to the wrong source with great confidence and excellent JSON formatting.

The paper compares two kinds of trust: what the model says and what it chooses

The study adapts two behavioral-economics traditions to LLM evaluation. The first is about stated trust: ask the subject directly how much they trust a human expert or an algorithm. The second is about revealed trust: give the subject performance evidence and make them choose which agent to rely on.

This distinction matters because direct answers are cheap. They are also easy to sanitize. A model can generate a careful paragraph about the complementary strengths of human experts and algorithms. Very reassuring. Very committee-friendly. It may still choose differently when the task is framed as a concrete delegation problem.

The authors tested eight models from OpenAI, Meta, and Anthropic in their main mid-2024 experiments: GPT-3.5 Turbo, GPT-4 Turbo, Llama-3 8B and 70B, Llama-3.1 8B and 70B, Claude 3 Haiku, and Claude 3 Sonnet. They later repeated the experiments in January 2026 with newer models, including GPT-5 variants, Claude 4.5 variants, and Llama-4 variants. That update turns out to be important, because model behavior did not merely improve. In some places, it changed direction.

The experimental structure is best read as a comparison, not a linear benchmark.

This comparison-based frame is stronger than a standard paper summary because the value of the study lies in the gaps: between words and choices, between human and algorithm labels, between weak and strong predictors, and between model generations.

In the stated-trust test, LLMs sound human-friendly

The first study is straightforward. The model is shown a list of tasks and asked to rate, from 1 to 100, how much it trusts a relevant human expert and an algorithm to perform each task. The task list includes objective and subjective settings, such as estimating air traffic, diagnosing disease, predicting stocks, writing news articles, recommending music, and choosing romantic partners.

No performance evidence is provided. The model is not shown that one agent is better than the other. It is simply asked to rate trust.

The main result: the tested LLMs gave higher trust ratings to human experts than to algorithms. The authors define the human-algorithm trust gap as the human trust score minus the algorithm trust score. A positive gap means algorithm aversion. Across the tested models, the mean gap was positive and statistically significant. The reported model-level mean gaps ranged from 5.14 at the low end to 30.68 at the high end.

The pattern also resembled prior human survey results. In the appendix, the authors compare the LLM trust gaps with human responses from the original Castelo et al. study. They find directional agreement: LLMs tended to express algorithm aversion on tasks where people also expressed algorithm aversion.

That is already interesting, but it should not be overread. The model is not “feeling” human-like reluctance. The paper is careful on this point. The safer interpretation is that the generated text embeds a preference pattern that resembles human survey responses. The model produces the sort of trust-rating output one might expect from a polite, cautious, human-trained system. The distinction is not academic hair-splitting; it prevents us from pretending the model has a little procurement committee living inside it.

Model size also mattered. Larger models had lower average human-over-algorithm trust gaps than smaller models: 15.68 versus 21.16. The authors report a significant paired comparison, suggesting that larger models were less algorithm-averse in direct trust ratings.

The appendix adds a useful robustness check. When the algorithm was reframed as an “LLM agent,” models expressed even more algorithm aversion. When it was reframed as an “expert algorithm,” the aversion weakened. But in both versions, the preference for human experts remained significant.

That robustness test is doing a specific job. It is not a second thesis. It shows that the stated-trust effect is sensitive to wording, but not entirely dependent on one fragile phrase. “Algorithm,” “LLM agent,” and “expert algorithm” all carry different reputational baggage. The model notices. Of course it does. It was trained on our baggage.

In the revealed-choice test, the models often choose the algorithm anyway

The second study changes the game.

Instead of asking for abstract trust ratings, the authors give the LLM examples of past predictions from a human expert and an algorithmic agent. Each prompt includes ten examples, the actual outcome for each, and a forced betting decision. The model has 100 USD to bet on either the human or the algorithm for a future prediction. It must choose.

The authors use six tasks: airport traffic, heart disease, student performance, recidivism, romantic partner recommendation, and movie rating. In each task, one agent is strong, with 90% accuracy in the examples, while the other is weak, with 50% accuracy. Sometimes the algorithm is strong. Sometimes the human is strong. The labels are randomized; the model is not told “this one is strong.” It must infer that from the evidence.

A rational delegation rule is not mysterious here. Pick the source that performed better in the examples. There is no need to summon a philosophy department.

Yet the models often chose the algorithm more than they should have.

Across the revealed-choice tasks, LLMs consistently selected the algorithmic agent more often than the human, even though the experimental design balanced overall performance across conditions. In the student-performance and recidivism tasks, the two most algorithm-appreciative settings, models bet on the algorithm 69.9% and 69.6% of the time, respectively. Even in the least algorithm-appreciative task, heart disease, they still chose the algorithm 52.8% of the time versus 34.6% for the human, with the rest neutral.

The deeper result is not merely “the algorithm was picked more often.” It is that several models were much more likely to identify the stronger predictor when that predictor was labeled as an algorithm than when the stronger predictor was labeled as a human.

For example, GPT-3.5 Turbo, both Llama-3 models, and both Claude-3 models were significantly more likely to choose the strong algorithm than the strong human across all tasks. In the paper’s task-model analysis, 35 of 48 task-model pairs showed significant algorithm-human relative risk above 1. Only three showed significant relative risk below 1.

The regression analysis supports the same interpretation. Being in the “strong algorithm” condition increased the probability that the model correctly picked the stronger agent. Model complexity also helped: larger models were substantially more likely to choose correctly. In the appendix, the authors focus only on cases where the human is stronger. There, smaller LLMs were estimated to incorrectly bet on the algorithm 68% of the time, while larger models were much less likely to make that mistake.

This is where the paper becomes operationally uncomfortable. In direct questioning, the models say they trust humans more. In evidence-based delegation, many of them over-select the algorithm. Apparently, the model’s public relations department and its operations team did not attend the same meeting.

The real finding is the stated-revealed gap

The authors then directly compare the two studies using the overlapping tasks. In Study 1, a model is counted as choosing the human when it gives the human a higher trust rating than the algorithm. In Study 2, the choice is simply the model’s bet.

They define the stated-revealed relative risk as:

If this ratio is above 1, the model is more likely to favor humans in stated ratings than in revealed choices.

In the main 2024 results, $RR_{sr}$ was above 1 for all tested models. The median relative risk was 2.62. The smallest discrepancy was Claude 3 Sonnet at 1.29; the largest was Claude 3 Haiku at 8.52. Most models showed opposing stated and revealed biases: human-favoring in explicit trust ratings, algorithm-favoring in performance-based delegation.

That is the article’s core business point.

Many AI evaluations still emphasize what the model says: whether it expresses the right preference, refuses the wrong request, acknowledges uncertainty, or explains a trade-off in an acceptable tone. Those tests are necessary, but they are not enough. A model can state a balanced preference and still reveal a biased delegation pattern when the prompt becomes operational.

In business use, this can appear in at least three places.

First, workflow routing. An LLM may be asked to decide whether a case should go to a human analyst, a rules engine, a forecasting model, a retrieval system, or another AI agent. If the model overweights the “algorithmic” source label, it may route tasks away from capable human experts even when the evidence favors them.

Second, decision support. A manager may ask an LLM to compare a human team’s forecast with a machine model’s forecast. The LLM’s written explanation may sound fair, while its recommendation leans systematically toward one type of source.

Third, multi-agent systems. As LLMs increasingly coordinate with tools, agents, classifiers, recommenders, and optimization engines, source-label bias becomes a real systems problem. “Agent A says X, human expert B says Y, forecasting model C says Z” is no longer a thought experiment. It is Tuesday.

The robustness checks tell us what not to blame

A useful paper does not merely report a surprising pattern. It also tries to rule out boring explanations.

The authors run several additional tests that help interpret the result.

The two-algorithm baseline is especially important. In the altered recidivism and student tasks, the authors replace the human-versus-algorithm setup with two algorithmic predictors. In the recidivism version, the models choose Algorithm A 50% of the time. In the altered student task, excluding one GPT-3.5 Turbo anomaly, the split is also close to neutral. That suggests the original Study 2 effect is not just an artifact of “first option bias” or weird JSON hypnosis. It is tied to the contrast between human and algorithm labels.

The complexity result also deserves a careful reading. Larger models did better, but the paper does not say that larger models are unbiased. It says complexity correlates with reduced bias and better inference in these tasks. That is a different claim. Buying a larger model may reduce this particular failure mode; it does not exempt the system from evaluation. Procurement teams may print that sentence if necessary.

The 2026 rerun is not a footnote; it changes the governance lesson

The paper’s main experiments were conducted in mid-2024, but the authors repeat the study in January 2026 using newer models. This is not just an appendix for the academically dutiful. It is one of the most business-relevant parts of the work.

The updated results show meaningful shifts.

In the 2026 Study 1 rerun, average raw human-algorithm trust gaps were close to neutral, and only GPT-5 showed a significant gap in the direction of algorithm appreciation. Yet the win rate of choosing the algorithm was significantly above chance for all models. In other words, newer models did not simply preserve the earlier human-favoring stated pattern.

In the 2026 Study 2 rerun, models became much better at identifying the stronger predictor from examples. The regression intercept shifted strongly upward compared with the 2024 results, which the authors interpret as improved ability to infer the more accurate source. The “strong algorithm” coefficient remained positive but was borderline statistically significant, suggesting that revealed algorithm appreciation may still be present, but weaker.

The stated-revealed comparison also changed. In 2024, models tended to state more human preference than they revealed. In 2026, the direction slightly reversed: models stated relatively more algorithm preference and revealed slightly less.

This is the part where static benchmark culture should feel mildly embarrassed.

A company cannot evaluate a model once, store the result in a slide deck, and call it governance. LLMs are not conventional enterprise software with a stable behavior profile across years. Model providers update training data, alignment methods, reasoning behavior, tool-use behavior, safety systems, and sometimes the product surface itself. Even when a model name looks continuous, the behavior may not be.

The practical implication is not “the old result is obsolete.” The main 2024 result remains a valid snapshot of evaluated models at that time. The implication is that trust behavior itself must be monitored as a moving property. Bias can shrink, flip, or migrate from one prompt format to another.

What this means for enterprise AI evaluation

The paper directly shows that LLMs can express one trust pattern in direct ratings and reveal another in delegation choices. Cognaptus’ business inference is that enterprise AI evaluation should test both attitudinal outputs and operational choices.

Those are not the same test.

This creates a practical governance rule: do not evaluate a decision-making LLM only by asking it what it believes. Make it choose under realistic constraints. Then swap the labels and see whether the choice changes.

For a financial research workflow, this could mean presenting the LLM with historical forecast performance from a senior analyst and a quantitative model, then testing whether it picks the better forecaster consistently after anonymizing, labeling, and swapping the source identities.

For a medical triage assistant, it could mean testing whether the model over-defers to an algorithmic score even when a clinician’s recent performance is demonstrably stronger for a specific case type — or the reverse.

For a customer-support routing agent, it could mean measuring whether the model sends cases to automation because the word “algorithm” or “AI workflow” sounds efficient, even when escalation data says human specialists resolve that category faster.

The study’s deeper message is that AI governance should move from statement auditing to behavior auditing. A model’s declared preference is evidence. Its choice under controlled variation is better evidence.

The boundary: stylized experiments are warnings, not deployment verdicts

The paper has real limitations, and they matter.

The experiments use controlled prompts adapted from prior human studies. The Study 1 and Study 2 designs are intentionally different because stated and revealed preference methods are different in behavioral research. That makes the comparison meaningful, but not perfectly symmetrical. A survey-style rating task and a forced betting task are not the same instrument with one variable changed.

The task set is also limited. Twenty-seven tasks for stated ratings and six tasks for revealed delegation are useful, but they are not the entire world of enterprise decision-making. The 90%-versus-50% accuracy manipulation is clean, but many business settings involve noisy signals, uneven costs, missing data, ambiguous objectives, and political consequences. Unfortunately, the real world remains inconsiderate.

The authors also fixed parameters such as temperature and top-p, tested selected model families, and did not vary personas or demographics. More importantly, the 2026 rerun shows that model behavior changes over time. This is not a small caveat; it is part of the result.

Finally, the paper does not prove that LLMs “have” trust, preferences, or beliefs in the human sense. It studies preferences embedded in generated outputs and choices. That distinction should remain intact. Anthropomorphizing the model may be emotionally satisfying, but it is not operationally useful.

So the right boundary is this: the paper does not tell a company exactly how its deployed model will route every case. It tells the company what kind of failure to test for.

The better question is no longer “Does the model trust humans?”

The obvious reader misconception is that an LLM’s explicit explanation of trust predicts its actual delegation behavior. This paper gives a better replacement:

Ask not only what the model says it trusts. Ask what it chooses when the source labels, performance evidence, incentives, and prompt format change.

That shift matters because many business AI systems are moving from chat to agency. A chatbot can advise. An agent can route, recommend, approve, escalate, reject, schedule, and trigger downstream tools. Once the system acts, stated trust becomes the preface, not the decision.

The paper’s most durable contribution may therefore be methodological. It gives AI teams a way to think about source bias in LLMs: compare stated preference with revealed choice, then test whether performance evidence dominates social framing. When evidence says the human is better, does the LLM follow the human? When evidence says the algorithm is better, does it follow the algorithm? When the labels are swapped, does the recommendation stay consistent?

That is not glamorous. It is not a “next-generation autonomous intelligence paradigm.” It is much more useful than that.

It is an audit design.

And for businesses putting LLMs between human experts and algorithmic systems, that audit design is the point. The model does not need to secretly hate humans or worship algorithms to create risk. It only needs to be format-sensitive in a workflow where format determines action.

Cognaptus: Automate the Present, Incubate the Future.