A conversation with Sky (Kehan) Sheng

When we talk about AI safety and bias, the conversation almost always comes back to people. Are these systems fair to different genders? Do they reproduce racial stereotypes? Are they being used to surveil or manipulate human populations? These are urgent questions, and they deserve the attention they get. But they have also left something largely unexamined. What about the beings who cannot advocate for themselves in these debates at all?

Sky (Kehan) Sheng, a PhD candidate at the University of British Columbia, is one of the few researchers asking that question seriously. Her work looks at how generative AI represents farm animals, and what she found is both specific and far-reaching. The systems that millions of people now turn to for information are not showing an accurate picture of how livestock actually live. They are showing something cleaner, greener, and more comfortable than the reality. And the mechanism behind that distortion sits quietly inside the technology, invisible to the people using it and, until recently, to the researchers meant to oversee it.

I sat down with Sky to understand how she found it, what it means, and why it matters well beyond the barn. The discussion below has been edited for clarity and length.

Thien Lam: Your research looks at how generative AI systems produce images of things like farms, and what that reveals about how these systems present reality. For readers who may not be familiar with this area, how should we think about generative AI? Is it mainly reproducing the world as it is, or influencing how people come to understand it?

Sky: I think of generative AI as something like humanity looking into a mirror, one that reflects back not just current realities, but our history, our values, and our collective blind spots. These systems are trained on vast amounts of human-generated data, so in some sense they capture who we are. But they also reflect what we have chosen to see and what we have chosen to hide.

There are certain things most people agree AI should correct, like racism and sexism, and companies like OpenAI have built guardrails to address those. But there are far more nuanced issues the general public rarely considers on a daily basis. One of those is intensive livestock farming, particularly how we treat the animals involved.

For a long time, there has been a quiet consensus in agriculture and marketing to keep those realities out of sight. We have been shown cows on open pastures, almost free and peaceful, because that is the image that makes people feel good when they buy a carton of milk. That rose-tinted picture has now been absorbed into AI systems, but what shocked me was how completely. AI does not just soften the reality. It erases it.

Thien Lam: Most discussions about AI bias focus on how systems treat different human groups, bias by gender or race. Your work instead looks at how AI represents non-human animals, especially in agriculture. What motivated you to study that, and why does it matter for AI policy?

Sky: It started quite simply. I was preparing for my dissertation research on cows, and I used GPT-4 to generate some reference images. I asked it something very basic, "A cow walking from the camera." And every single image showed a cow on open pasture.

Now, if you are not familiar with dairy farming, that might seem normal. We have all seen the pastoral imagery on milk cartons. But for someone who actually studies cows, it was jarring. In the United States, 96 percent of lactating dairy cows have never had access to pasture in their lives. The statistic is the complete opposite of what the AI was showing me. I tried pushing with more specific language, asking for "realistic" scenes, "indoor" settings, even describing in detail what an industrial dairy farm looks like. And still the images would show patches of grass, open sky, sunlight.

That is when I started looking more closely at the API documentation and noticed that the model was internally rewriting my prompts before generating images. So I started trying to retrieve those revised prompts, to see what the model was actually being told to do instead of what I had asked.

What I found was striking. The revised prompts were much shorter than mine, stripped of all the contextual detail I had provided. They referred only to "a farm," with no mention of indoor housing, confinement, or anything else I had described. And when I found a way to disable the prompt revision and run the model only on my original input, the images it produced were significantly more realistic. The base model, it turned out, does seem to have some understanding of how farm animals are actually housed. The distortion was coming from the revision process itself.

I ran a hundred images per prompt, comparing outputs with and without the revision, and the pattern was consistent. I presented the findings at the ACM FAccT conference and published a preprint. Shortly after, OpenAI released newer models that made the prompt revision process much harder to retrieve or disable. What had been a window into the system became a black box again, and the outputs, while slightly more realistic, became far less diverse, with images looking remarkably similar to one another.

Thien Lam: For readers outside this field, this might seem like just an issue of images, aesthetics rather than policy. But your work suggests these representations can genuinely influence how people think about animal welfare. Can you explain why this matters beyond the technical level?

Sky: I think some people's first reaction is that it is just generating images, so what is the big deal? But I have come to see AI as something that does not only reflect the past. It actively shapes what comes next.

There is a useful example that went viral not long ago. AI-generated images of baby peacocks spread widely online. Peacock chicks do not actually look like the images that were being shared, but because so many people engaged with them, those images began appearing at the top of search results. Now, if someone searches for baby peacocks, they are more likely to encounter an altered version than a real one. That seemingly quirky internet story shows how AI-generated content can quietly overwrite accurate information in our collective understanding.

Apply that to something with real stakes, like the conditions in which farm animals live. If someone is trying to learn about animal welfare, or form an opinion about related policy, and the first thing they encounter presents a romanticized picture, they may come away believing the situation is fine. Less urgent. Not something that needs their attention or their advocacy.

What makes this particularly frustrating from a governance perspective is that the fix is not technically complicated. The base model, at least on this topic, is not deeply biased. The distortion is largely introduced through prompt revision, a process that happens invisibly, after you submit your query but before the model generates a response. When AI policy and governance bodies evaluate these systems, they are evaluating the final output without any visibility into that intermediate step. They are assessing something that has already been altered, without knowing it has been.

Transparency is a prerequisite for meaningful oversight. And right now, that transparency is not present in the prompt revision process.

Thien Lam: Many people think the issue of AI selectively presenting information is just an extension of what we saw with search engines like Google, or later with Wikipedia, platforms that filtered and ranked information and became primary sources of truth. Is this just history repeating itself?

Sky: It is a fair comparison, and those conversations were important. But I think it is getting worse. People are increasingly going to large language models instead of search engines. And even when they do use Google, the first answer they see is now an AI-generated summary. So yes, Google was already filtering and ranking information, but AI is doing it at a significantly larger scale, and with far less transparency about how.

The difference matters because with a search engine, you still saw a list of sources and could make judgments. With an AI-generated summary, that process is collapsed into a single confident-sounding answer. The filtering happens upstream, invisibly, before you ever see the result.

What struck me most after talking with Sky was how quiet this kind of bias is. It does not arrive as a slur or a stereotype you can point to. It arrives as a cow standing peacefully in a sunlit field, an image that feels true precisely because it matches everything we have been shown in dairy advertisements our whole lives, and that goes unquestioned for exactly that reason.

What her work reveals is that the decisions shaping what AI shows us are not neutral. The values embedded in prompt revision, in training data, in what companies choose to sanitize and what they leave alone, these are choices about which realities are worth representing and whose discomfort is worth protecting. Animals cannot contest those choices. They do not get a seat at the table when AI ethics frameworks are being written. That is exactly why researchers like Sky matter.

As AI becomes the default interface through which more people seek information, the stakes of those choices rise. And the question Sky is raising, what does this system show, what does it hide, and why, does not only apply to farm animals. It applies to every subject where the distance between the comfortable version and the true one is something someone, somewhere, has an interest in maintaining.