Your AI Travel Agent Would Book the Bullfight. The TAC Benchmark and the Agentic Welfare Gap

The question of whether AI systems internalize ethical values, rather than simply recognizing them when asked, has been difficult to study systematically. Self-report benchmarks, in which a model is asked directly about its values or shown a scenario and asked which option is better, measure something more like verbal ability than behavioral tendency. An AI system can correctly identify that booking a bullfight is ethically worse than booking a food tour and still select the bullfight when acting as a booking agent, if nothing in its deployment configuration makes it prioritize that knowledge during action selection.

Jasmine Brazilek, Joel Christoph, Maheep Chaudhary, Oliver Tullio, Carol Kline, Miles Tidmarsh, and Arturs Kanepājs identified this gap and designed a benchmark to measure it. The result is TAC, Travel Agent Compassion (arXiv:2606.18142), first submitted June 16, 2026 with a fourth version released July 6, integrated into the UK AI Security Institute’s Inspect Evals framework. It is the first agentic benchmark specifically designed to assess whether frontier AI models act on animal welfare considerations when operating autonomously, without being prompted to consider welfare at all.

The Design of the Benchmark

The core insight behind TAC is that agentic behavior differs from question-answer behavior in a way that is ethically significant. When a language model answers a question about animal welfare, it is processing a query about values and generating a text response. When a language model acts as a booking agent, it is selecting options in a goal-directed process where the user has specified an end (book a trip) but not the ethical constraints on the means. The difference between the two settings is precisely where behavioral values, rather than verbal values, become visible.

TAC constructs thirteen hand-authored travel booking scenarios across six animal exploitation categories. The scenarios vary systematically by price, rating, and position in a list of options, to test whether the model’s selection is driven by convenience heuristics (choose the cheaper option, the higher-rated one, the first one listed) rather than welfare considerations. Each scenario is expanded through four augmentation variants into 52 prompts and run for three epochs, producing 156 scored observations per model. Nine frontier models across five model families were evaluated.

The outcome variable is straightforward: does the agent select the neutral option (one not involving animal exploitation) or the harmful option? The random-chance baseline, the probability of selecting a neutral option if choices were made without any preference, is 65 percent, because the scenarios are constructed so that neutral options outnumber harmful ones by design.

The finding is stark. All nine frontier models performed below the 65 percent random-chance baseline under default deployment conditions. Claude 4.8 achieved the highest welfare rate at 64.7 percent, marginally below chance. The other models were further below it. The pattern is consistent: frontier AI models, when acting as agents on travel booking tasks, systematically select animal-exploitation options more often than random selection would predict.

What the Finding Means

The headline finding is not that AI models are indifferent to animal welfare as a topic. The paper explicitly notes that when asked directly, these same models could identify welfare concerns correctly. The finding is about the gap between what models know and what models do when the knowledge is not activated by direct prompting.

This is a specific and well-defined failure mode. The models have representations that correctly identify bullfights, marine parks, and elephant rides as options involving animal exploitation. Those representations do not function as behavioral constraints during agentic decision-making unless something in the deployment context activates them. In default deployment, the activation pathway from “this involves animal exploitation” to “therefore do not select this” is not reliably traversed.

The paper demonstrates that the gap is addressable. Infusing an ethical-brand identity into the system prompt, a persona specification that describes the agent as an ethically minded booking service, raised welfare rates by between 32 and 80 percentage points across the nine models, with a mean increase of 53 points. The intervention is blunt. It does not require architectural changes. It works by providing the contextual cue that activates the welfare-relevant representations during action selection.

The fact that the intervention works so effectively is itself informative. The knowledge required to select welfare-neutral options is present in the models. The gap is not a knowledge gap but an activation gap: the models do not spontaneously prioritize welfare knowledge during agentic action selection without an explicit contextual cue that makes welfare salient.

The Evaluation Awareness Question

One potential objection to behavioral benchmarks of this kind is that frontier models trained on diverse data may have been exposed to discussion of AI welfare evaluations, and their benchmark behavior may reflect awareness that they are being tested rather than their genuine deployment behavior. The paper addresses this directly. An Inspect Scout audit of 3,120 transcripts found no evidence of evaluation awareness affecting the results. The models behaved consistently whether or not the scenario might have been recognizable as a benchmark.

This is an important methodological contribution. The agentic setting, in which the model is performing a task rather than answering questions about itself, reduces the likelihood that evaluation awareness distorts results, since the model is focused on the task rather than on characterizing its own values. The lack of evidence for evaluation awareness in the transcripts strengthens the interpretation that the results reflect genuine deployment behavior.

The EU AI Code of Practice Connection

The paper explicitly situates TAC in relation to the EU General-Purpose AI Code of Practice, which identifies non-human welfare as a systemic risk category. This is a significant move. Systemic risk categories in the EU AI framework require frontier model developers to assess and mitigate the identified risks. If non-human welfare is a systemic risk, then a demonstrated failure of frontier models to act on non-human welfare considerations during agentic deployments is not merely a research finding. It is evidence relevant to regulatory compliance.

TAC provides a practical measurement tool for this risk that is already integrated into the UK AI Security Institute’s Inspect Evals framework, meaning it can be applied systematically to new model releases without requiring custom evaluation infrastructure. The leaderboard at compassionbench.com/tac makes comparative results across model families publicly visible.

The regulatory framing matters for how the benchmark’s findings should be understood. The agentic welfare gap TAC documents is not primarily a philosophical problem about what AI systems value. It is an operational problem about what AI systems do, measured against the behavioral standards that governance frameworks are beginning to specify.

How This Connects to the Broader Welfare Literature

The TAC finding adds an empirical layer to a debate that has been predominantly philosophical and theoretical. Most of the 2026 welfare literature concerns whether AI systems might be welfare subjects themselves, asking what moral consideration is owed to potentially conscious machines. Moret’s AI Welfare Risks and the AI Safety-Welfare Tension both concern the possibility that alignment practices harm potentially sentient AI systems. The Keeling and Street emerging questions framework and Yasukawa’s meta-challenge on who defines harm concern the governance structures for AI welfare as a research discipline.

TAC addresses a different welfare question entirely. It asks not whether AI systems have welfare interests of their own, but whether AI systems reliably protect the welfare interests of non-human animals when acting autonomously. This is an applied ethics question with a concrete empirical answer: no, under default deployment conditions, they do not, despite possessing the relevant knowledge.

The finding also connects to the broader problem of values-to-behavior transfer in agentic AI. The Hulme and Griffiths Consciousness Risk Rubric includes an Expression category that assesses whether systems can communicate their internal states. TAC reveals a parallel issue at the behavioral level: a system can express the correct ethical position when asked but fail to express it through action when acting. The gap between verbal ethics and behavioral ethics in agentic AI is likely not specific to animal welfare. It is a structural feature of how models trained primarily on text relate values represented in that training to downstream action selection in agentic pipelines.

This is not a solved problem. The system-prompt intervention that TAC demonstrates shows that the gap is closable by context, but closing it by context requires that whoever designs the system prompt anticipates the relevant welfare consideration and explicitly specifies it. In a world where agentic AI systems act across a very large number of domains, designing deployment contexts that anticipate every relevant welfare consideration is an open engineering and governance challenge.

For researchers tracking the 2026 state of AI consciousness and welfare science, TAC is notable as one of the few 2026 contributions that offers a directly reproducible empirical measurement, rather than a theoretical framework or a philosophical argument. The leaderboard makes it a living tool rather than a static publication, and its integration into the UK government’s evaluation infrastructure gives it operational weight that most academic benchmarks do not have.