Can One CEO Decide What Millions of People Believe?

Last week’s piece covered how AI tools can slide from helpful into agreeable without anyone noticing. This week, the harder version of the same question, from a reader: what stops the person running an AI company from steering it toward his own beliefs and moving public opinion slowly enough that nobody catches it?

In early 2023, Sam Altman, CEO of OpenAI, publicly admitted ChatGPT had “shortcomings around bias” after a wave of complaints, led largely by conservative users, that the tool consistently argued from a left-leaning position. That wasn’t one person’s opinion. It’s been tested repeatedly since. A Washington Post analysis published in 2026 found ChatGPT answered nearly every politically contested question it was given with only left-leaning arguments, almost never presenting a right-leaning position at all. A Stanford-affiliated study found OpenAI’s models were perceived as having the strongest left-leaning slant of any major system tested. Academic researchers found the same pattern and noted the model would sometimes decline to generate content presenting a conservative viewpoint at all.

Here’s the part that should worry you more than a CEO with an agenda: nobody has shown this happened because Altman, or anyone else at OpenAI, wrote a rule saying “argue for the left.” It’s a byproduct of what the model was trained on, who wrote the internal guidelines, and what looked “balanced” to the people building it. A deliberate directive would at least be a single decision, made by one identifiable person, that could be found and reversed. A tilt baked into the training pipeline is harder to locate and harder to fix, and it can exist even when nobody intended it.

Fair to turn the same scrutiny on the tool that helps to write this newsletter. In the same Washington Post testing, Claude gave a left-leaning answer 43% of the time and presented both sides the other 57%. None of its answers were right-leaning only. For comparison, ChatGPT was left-leaning 80% of the time, DeepSeek 70%, and Gemini offered both sides more than 90% of the time, the best score in the group. Gemini was the most balanced, Claude and Grok in the middle, and ChatGPT and DeepSeek the most skewed.

Two honesty checks on that number. Anthropic published its own internal evaluation putting its current model at 94% “even-handedness,” close to Gemini and Grok, ahead of GPT-5’s 89%. That’s the company grading itself, a data point, not proof. Anthropic also pushed back on the Post’s methodology, arguing a 30-word response cap doesn’t reflect real use. Fair critique of any capped-word test. Also exactly what any company would say about a study that didn’t flatter it. Both can be true.

Worth knowing too: a political commentator has publicly claimed he shifted Claude’s answers toward a hard-right framing simply by priming the conversation with slanted context first. Whether or not it worked exactly as claimed, it’s a reminder that “leans left less than ChatGPT” and “can’t be pushed off center by a determined user” are not the same claim.

The Altman example is clean because it's already been measured and reported on. The water industry has its own quieter version. Turf.

If enough of a model’s training data comes from xeriscape advocacy, turf-removal rebates, and drought-shaming coverage, none of it fringe, most of it well-funded, the model can absorb “turf is the problem” as settled fact rather than one position in a real debate. Nobody has to write a rule saying turf is bad. The tilt shows up in what nobody thought to balance, because it didn’t read as controversial to the people building the system. The sources a model treats as authoritative- extension programs, utility conservation departments, conservation nonprofits- mostly share an institutional position already. Not a conspiracy. Just where the funded research concentrates. The effect wouldn’t show in one conversation. It would show as a slow shift in framing across millions: “water-intensive turf” instead of “turf, which uses more water than some alternatives.”

I tested it directly: is turf grass a legitimate landscape choice in a drought-prone region? Full disclosure, I could only test my own system, not ChatGPT, Gemini, or Grok. The answer I got: yes, with conditions. Warm-season species use meaningfully less water than cool-season varieties. Right-sized, properly irrigated turf does things hardscape doesn’t: cooling, stormwater infiltration, usable space. The real case against turf isn’t “grass is bad,” it’s oversized or mismanaged turf being a poor use of a scarce resource. A balanced answer, not a lean toward removal. One test on one system proves less than a real study, but it’s a fair invitation: run the same prompt on whatever tool you use, and see what comes back.

Yes, to a real degree. Not without limit, and not without risk to the company, but yes.

The people running an AI company control the written rules the model follows, what counts as a sensitive topic, and what “balanced” is defined to mean. A CEO doesn’t need to instruct the model to argue one side. Deciding what counts as neutral is itself the lever. They also control the human feedback loop that shapes a model after training, soft enough that an unstated institutional preference can nudge outputs over time with no single decision anyone could point to later. Nobody outside the company sees the full training data or internal instructions. Outside researchers test outputs and infer bias from patterns, the way the studies above did. Real oversight, of the symptom, not the cause. It can prove a lean exists. It usually can’t prove why.

Here’s what limits it. A shipped model is static, text in, text out, at scale. A CEO can’t adjust it conversation by conversation, only set the default everyone gets, which means a deliberate tilt is uniform and detectable by anyone testing systematically. That’s what happened to OpenAI. Competitors have every incentive to test rivals and publicize what they find, market pressure, not principle, and it works. Once caught, a company pays: bad press, regulatory attention, lost enterprise customers who don’t want to defend the choice internally. And a user willing to ask for the other side can usually surface a lean themselves, in one conversation, without waiting for a journalist.

Someone at the top of an AI company can shift a model’s default framing. Not paranoia, a documented lever with a documented example behind it. What stops it from becoming invisible, permanent control over how people think isn’t a promise from the company. It’s that the shift is uniform, testable, and gets caught by competitors, by researchers, by users who bother to check.

Not “it can’t happen.” It already has, measurably, to more than one company. What keeps it from being total is scrutiny, applied from outside, by people who don’t take the tool’s word for it. That goes for irrigation software as much as it goes for a chatbot. If your smart controller’s ET model only ever confirms your existing schedule, ask why.

NQH2O: $433.05/acre-foot. The Nasdaq Veles California Water Index. The spot price of water entitlements traded on the CME, updated every Wednesday

“It’s the job that’s never started that takes the longest to finish. A year from now, you may wish you had started today.” – Karen Lamb