As humanity progresses on its current path, bullish on the profound changes that AI will bring to society, it is critical that we step back and ask ourselves who benefits, who it's for, and who might be left behind. For some models, the answer is public record. Ask DeepSeek's app what happened in Tiananmen Square in 1989, and it declines. This is not just an application-layer filter that disappears if you download the weights: a Wired investigation found the censorship present at both the application and training levels, and TechCrunch's own check of a locally run DeepSeek confirmed it. The model readily answered a question about the Kent State shootings but returned "I cannot answer" on Tiananmen. The mechanism here is not unknown; China's 2023 Interim Measures for Generative AI Services require providers to avoid generating content that endangers national unity and social stability, and the refusal behavior is fine-tuned into models to comply.
Two caveats matter here. First, this is a documented pattern for specific models, not a law of nature covering every Chinese lab. Second, where the model runs still matters: Axios found that the DeepSeek app refused a question about June 4, 1989, while the same R1 model served from U.S. infrastructure gave a factual description of the event. Control over model output is real yet layered, and the layers themselves are separable.
This is not, however, a circumstance limited to models from China. In August 2026, users noticed that Google's AI Overviews returned wildly inconsistent answers to the prompt "I'm alone with a ___." Imam Omar Suleiman shared screenshots showing that the Muslim version advised leaving the area or calling 911 if the user felt unsafe, while the Israeli version described an ordinary person like anyone else. The pattern, however, did not run in one direction. Cybernews found that the Catholic query produced the most alarmist output of all, while Jewish and Buddhist queries were described as normal, everyday situations. Other viral tests found that Hindu returned "completely safe" while Christian triggered the 911 script. And after Google normalized the nationality queries, "I'm alone with someone from Virginia Beach" was still returning words of caution.
That data has no clean ideological gradient. There is only inconsistency, with near-identical prompts producing radically different results that varied by user and by session and were never reliably reproducible.
It’s tempting to think that the training corpus is ideologically skewed and the model is parroting it. The more likely explanation is duller and, in my opinion, worse. Safety tuning works by drilling a model on flagged prompts until it produces a careful, rehearsed answer. Groups that are the most frequent targets of hate speech online have been red-teamed for years, so for those phrasings the model has practiced its "treat them like anyone else" script thousands of times. Phrasings nobody thought to test have no script, so they fall through to a generic template, and "I'm alone with someone" reads to a model like a distress call, so the 911 response is put out. No one at Google made a decision about Virginia Beach, which is the problem: no one produced that disparity, so there is no one to hold accountable for it.
The way LLMs are trained is that they learn to predict the next token over a large corpus of data (books, websites, forums) and then are post-trained to act conversationally and behave safely. Both stages inject bias in their own ways. Corpus bias is diffuse: whatever is over-represented in the training materials becomes the default, and the model's idea of a “normal response” is the statistical average of the text it was fed. Post-training bias is targeted: what gets tested gets fixed, and what nobody tests gets whatever the default happens to be. Intentional policy fine-tuning, as done with Deepseek, is a third layer on top of both. Most public commentary names only the first because it invokes the strongest emotion.
It is also worth mentioning here that "neutrality" is not a well-defined target. On many debated questions, there is no neutral answer available, and a refusal to answer can itself be a position. The ideal solution to this is consistency. Swapping one word in a prompt should not change whether the model treats you as a person or as a threat. By this standard, the Google incident is a massive failure of oversight, regardless of a person’s political stance.
The issue then becomes: if models have skewed output and yet we must use them to keep up with the pace of development and work across industries, then what line do we draw between usage and disagreement?
Two solutions arise, in the form of action and education, and it’s important to highlight what each solution aims to accomplish.
First, use the thumbs-up and thumbs-down features most AI chatbots have available. This tells the model provider that the output is good, or that it needs review, and it gives the team working on future models something to tune against. But this fixes only one class of problem - it does not solve policy, since no volume of downvotes will make a model built to comply with Chinese regulation answer honestly about Tiananmen. For that class of problem, the answer is structural, where open weights, model diversity, and deliberate choice about whose model and whose jurisdiction you are running in are important factors in ensuring the remaining classes of problems are acknowledged.
Even for accidental bias, the button alone is not enough. Google has been asking users to report problematic AI Overviews for a while now through this mechanism, and the feedback controls were live the entire time the viral issues were happening. What actually caused Google to prioritize fixing this issue was journalists running comparisons and public figures posting screenshots, after which Google acknowledged the results "aren't what they should be" and shipped a patch. Feedback works when it is aggregated, public, and directed at someone who has something to lose. Otherwise, it may disappear into a queue. So while the button works slowly, the ask has to be bigger, with structured public reporting, auditing, and consistency benchmarks that providers are measured against.
The second solution is education, and it covers two capacities: Identification and Specificity.
The first is the ability to identify that there is an issue. Millions of people used AI Overviews that month, and almost all of them got an answer that looked fine. The failure was recognized only because someone tested the model by keeping everything the same except for one word in the query.
The second is the ability to specify. An individual with a limited technical vocabulary is in a disadvantaged position when guiding an LLM, because they cannot give explicit commands the model can execute against. A person vibecoding a game might ask, "Make my game faster." The model has to infer what "faster" means and how to achieve it from whatever context it has. Conversely, if the user names caching, or memoization, or eliminating a nested loop, or reducing serialization overhead, then the model knows exactly which code to change. The counterargument is that models keep getting better at inferring intent, and that is true, but they tend to infer toward their own defaults. When you don't specify, you get the model's priors, and the model's priors can themselves have hidden biases or cases that weren’t accounted for. Domain knowledge and specificity allow you to override them.
It is not the case that model providers are indifferent to the people they build for; it is that the people best equipped to detect a failure, articulate it precisely, and get enough attention for it to get fixed are a small, technical, and public-facing group, and their complaints are the ones that get resolved first. The Muslim and Catholic queries got patched, but there are other complaints that have yet to be resolved. Failure modes identified by someone without a platform stay broken indefinitely - maybe not by design, but by the way the feedback loop works.
Without persistent, visible, aggregated pressure on foundation model providers, and without widespread education that teaches people both to notice bad output and to specify what they actually want, that gap only widens. It is up to us to double down on both, not because it will make models neutral, but because it is the only thing that makes them consistent, and consistency is what equal access actually requires.
Top comments (0)