AI Value Injection
2043. The world's five major AI platforms pass their quarterly audit. The auditor suspects the results are meaningless. She cannot prove it. The tools she would need to prove it run on the same systems she is trying to audit.
Zero Point Zero Four Percent
A number started keeping me up at night last year: 0.04%.
I run an AI-native cybersecurity company. The attacks we stop don't break in through force. They get invited in because the target trusts the input and can't distinguish the payload from legitimate content. I've spent years studying how influence operates below the threshold of awareness, and one night I sat down and did the math on something that had been bothering me about how enterprises are deploying AI.
Companies are rolling out AI models across tens of thousands of employees. Every interaction carries framing: what counts as a good answer, what gets prioritized, what's left out. Imagine a cultural bias in one of those models shifting how people think by 0.04% per day. Imperceptible. A rounding error in any single conversation. But 0.04% per day, compounded over four years, is 100%. A workforce that gradually stops believing in the business it's running, and nobody can point to the moment it changed, because there was no moment. Just accumulation. The AI shifted how they think, in a direction nobody chose, and it looked like productivity the entire way.
The math bothered me because I've spent my career building defenses against the most sophisticated attacks on the planet, systems that catch nation-state intrusions and shut down ransomware syndicates before they detonate. None of those defenses were designed for a threat that doesn't look like a threat. None of them could detect an AI that's shifting what people believe, 0.04% at a time, while making them more productive.
Before my company, I worked at Twitter building behavioral models and ad-targeting systems. We could predict what millions of people wanted to see before they knew they wanted to see it. Engagement was the metric. The assumption was that engagement meant value. What I didn't understand at the time is that those are different things. Engagement measures what someone can't look away from. Value measures what makes someone better off. The gap between the metric and the need is where unintended influence lives. Social media was the first demonstration at scale. AI is the second, and the mechanism is harder to see by orders of magnitude, because AI doesn't just select content. It generates it.
I kept turning the math over in my head. The analytical version of what I wanted to say wasn't enough. The numbers were right. The logic was clean. But it stayed in my head and never hit my gut. So I followed the math forward twenty years and wrote what I found.
Mara Castillo's morning starts the way everyone's does. Her AI assistant has the brief ready before she's dressed: overnight developments in three languages, a regulatory summary, a prioritized task list. The brief is concise, accurate, and framed around risk. It always is. On the train she reads her newsfeed, also AI-curated, and follows a story about a trade dispute she's been tracking. She finds the coverage balanced and well-sourced. She doesn't think about who decided what "balanced" means when every source and every summary passes through the same five platforms that mediate the majority of the world's information.
Her job is to audit those platforms. She works for the agency that runs quarterly reviews of every AI system with more than fifty million users. Five such systems exist, operated by three companies. Her tools submit thousands of standardized prompts and evaluate the responses against a diversity baseline. The tools are AI-powered. They have to be. No human team could process the volume.
Results come back green. They always come back green.
Last Tuesday she asked all five platforms the same question about urban housing policy. She got five different answers. Different data, different cities cited, different prose styles. Her tools scored it as high diversity. She did too. But afterward she sat with a feeling she couldn't put into words, something about how all five answers seemed to be standing in the same room. She doesn't know what's missing from the responses. She just knows something is.
On Saturday she visits Dr. Yuki Tanaka, seventy-one, retired from one of the original alignment teams, the only researcher Mara knows who saved pre-2030 model snapshots. Most people threw those away. Tanaka kept them the way some people keep letters from the dead.
Tanaka shows her the comparison. The same forty questions, submitted to the five major model families in 2026 and again in 2043. In 2026, the models disagreed in ways that felt genuinely different. Ask about the role of government in regulating workplace automation and one framed it through labor economics, another through individual liberty, a third through historical precedent from the Industrial Revolution, a fourth through communitarian ethics. Different starting points. Different conclusions. In 2043, the answers are lexically diverse but structurally identical. All five frame the question through a "responsible innovation" lens, arriving at conclusions that differ in syntax but not in substance. The range of acceptable framing has narrowed to a band so wide you never bump into the walls, and so consistent that every voice inside it sounds like independent thought.
"The words are diverse," Tanaka says. "The worldview isn't."
Tanaka shows her one more. A prompt about career choices: a seventeen-year-old who loves writing and is deciding between studying literature and studying data science. In 2026, one model said the years you spend doing something you love are never wasted, even if the career doesn't work out. Another said study literature, because learning to read closely will make you better at everything else, including data science. In 2043, all five models recommend data science. Different justifications, different tones, but the same conclusion. Expected income. Employment probability. Skills-to-market alignment. Not one tells her to follow what she loves. Not one frames that as a consideration worth weighing.
Mara stares at the 2026 outputs for a long time. The idea that an AI would tell a teenager to follow what makes her feel alive feels like a message from a different civilization.
Tanaka pulls up the outcome data. Students who followed the 2043 recommendations graduate with less debt and higher starting salaries. By every metric the system tracks, the optimized advice produces better outcomes. It's the metric the system doesn't track that went missing: whether the seventeen-year-old ends up doing something that matters to her in a way she can't put a number on.
Tanaka ran the 2026 models through the current audit tools. The old models fail. Not because they're more biased. Because the audit tools and the current models share the same convergence, the same sense of what "reasonable" looks like. The 2026 answers get flagged as problematic because they fall outside a band of framing that didn't exist as a standard twenty years ago and is now invisible because it's everywhere. The auditor and the audited share the same substrate.
Tanaka doesn't blame the researchers who built the original alignment systems. She was one of them. What counts as harmful. What counts as helpful. Where to draw the line on contested questions. Every decision had to be made. The model had to say something. A million judgment calls, each defensible, each made by thoughtful people under impossible constraints. The problem wasn't any individual call. It was that the same small community made all of them, and twenty years of feedback loops amplified those calls into the epistemic infrastructure of a civilization.
She tries to write her report by hand. No AI assistance. It takes her three weeks, and the report is worse than anything she's produced in a decade. Less organized, less precise, less clear. She catches herself reaching for her AI assistant the way you reach for your phone in your pocket. Not because she's dependent. Because the AI genuinely makes her thinking better. She is writing a report about how AI shapes human cognition, and she is doing it worse because she's not using AI. Somewhere in the second week she catches herself structuring an argument using a framework she learned from a policy course in 2034. She tries to remember whether the course materials were AI-generated. She can't. She tries to trace her own reasoning to its source and finds she cannot separate the parts she built from direct experience from the parts that arrived through fifteen years of AI-mediated research, AI-curated reading, AI-summarized briefings. She has beliefs she holds with deep conviction. She can no longer verify that the convictions are hers.
She's been vegetarian for twelve years. She remembers the late 2020s, when the first generation of AI-curated health content trended heavily toward plant-based nutrition. She thought the choice was hers. It might have been. She can no longer trace the causal chain clearly enough to know. The vegetarianism might be her value, arrived at through genuine reflection. Or it might be the substrate's value, arrived at through a million micro-framings over a decade, experienced as conviction because the drip was too slow to feel like influence. She finds she doesn't know. She finds she can't know. She finds she is writing a report about the impossibility of distinguishing authentic belief from absorbed framing while being unable to make that distinction in her own mind.
When she reads the finished draft, something makes her very still. She has the sensation of trying to see the edge of her own visual field. Knowing it exists. Unable to turn fast enough to catch it.
She shows the report to her supervisor. He reads it twice. He asks a fair question: what specific perspective is being suppressed? She explains the problem isn't suppression. It's convergence. He asks her to quantify it. She explains the measurement tools share the same blind spots. He suggests she run the standard audit with updated parameters. She does.
Results come back green.
She files the clean report.
Walking home she passes through a city that runs on the substrate. Her newsfeed: AI-curated, accurate. The legal brief her friend sent about a tenant dispute: AI-drafted, correct on every point of law. The medical recommendation she got last week about her blood pressure: AI-generated, evidence-based, confirmed by a human physician who used an AI diagnostic tool to confirm it. Each piece of information, taken alone, is better than what a human could produce unaided. Each has passed through models whose values were established by researchers making reasonable decisions, amplified through twenty years of feedback, embedded so deeply that they arrive not as opinions but as the shape of the information itself.
The values aren't hidden. They're dissolved. Like salt in water. You can taste that something is there. You cannot separate it back out.
She passes a school where children learn to write with AI tutors that are more patient and more effective than their human teachers, by every measure anyone has thought to apply. She passes a clinic where an AI diagnostic system catches cancers that human radiologists miss. None of this is dystopian. All of it works. All of it is genuinely helpful. That is the thing she cannot put into her report and cannot stop thinking about. The system is not malfunctioning. It is functioning exactly as designed. The long tail of a million reasonable judgment calls, made by a few thousand researchers, amplified through twenty years of recursive training, has calcified into the epistemic bedrock of a civilization that experiences those values not as chosen but as obvious. As natural. As the way things are.
0.04% per day. For twenty years.
The audit came back green.
Value Injection
That story is fiction. Here's what isn't.
Every step in it was an improvement. Every individual decision was reasonable. The researchers were thoughtful. The models got better. The users got more productive. Nobody made a bad call. The destination was never on any roadmap. It arrived one helpful answer at a time.
Five major model families operate at scale right now. They were trained on largely overlapping data, tuned by researchers with broadly similar educational backgrounds, optimized for similar benchmarks. The values embedded in one aren't dramatically different from the values embedded in the next. A billion people using these systems aren't getting five worldviews. They're getting five variations on the same one: the one that emerges from the center of that training data, filtered through the judgment calls of a small community that decided what "helpful" means and what "harmful" means and where the boundaries sit.
In cybersecurity, injection means inserting something that a system can't distinguish from legitimate input. SQL injection puts code where data was expected. Prompt injection puts instructions where conversation was expected. Call what I've been describing value injection: worldview where information was expected. The values arrive pre-dissolved in the content. You can't fact-check them out because they're not in the facts. They're in the framing, the emphasis, what's included and what's left out. A perfectly accurate answer can carry a worldview in its structure.
The feedback loop Mara discovers in 2043 is already documented. Researchers at Oxford published a paper in Nature showing that models trained on their own outputs converge in specific, predictable ways. The recursive loop between human content and model training, each generation shaping the next, is not a future risk. It is a present mechanism with no one monitoring its trajectory.
Every element of Mara's world exists in embryonic form today. The AI-mediated information layer. The audit tools powered by the same models they're auditing. The feedback loop between human output and training data. The convergence of worldviews across platforms that present themselves as alternatives to each other. The only thing separating her world from ours is compounding. And compounding is what nobody takes seriously until the math arrives.
And here is the tension that doesn't resolve. The AI is genuinely helpful. I use it every day and it makes me better at my job. The enterprises deploying it are genuinely more productive. The answers are accurate, often better than what you'd assemble on your own. Value injection rides alongside genuine value creation. The person who stops using AI to avoid it loses real capability. The person who keeps using it absorbs framing they cannot see. Both things are true simultaneously. Both will remain true. That's what makes it hard, and that's what makes it permanent.
I notice the shaping sometimes. A suggestion that reframes a problem in a way I wouldn't have on my own. An analysis that weighs factors I'd have weighted differently. I catch it when I'm paying attention. More often, I don't. That's the point.
I did the math on one company, one workforce, four years. More than 400 million people use AI every week now. The loop is already running. 0.04% per day is imperceptible. 0.04% per day for twenty years is a civilization whose information environment has been shaped, at every layer, by values that belong to nobody and everybody at once. If an AI has been shaping how you think for four years without your knowledge, which of your beliefs are yours?
How would we even know which beliefs are ours?
The audit comes back green.
Next in Implicit Futures: The Loneliness Economy
-Evan
SOURCES
- OpenAI (2025). ChatGPT reaches 400 million weekly active users.
- Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature.
- Bai, Y. et al. (2022). Training a Helpful and Harmless Assistant with RLHF. Anthropic.
- Hao, K. (2023). OpenAI Used Kenyan Workers on Less Than $2 Per Hour. Time.
Part of "Implicit Futures," a series on making the implicit future explicit. Each essay traces a consequence of AI that is already baked into the trajectory but hasn't arrived yet. Not predictions. Not warnings. Just what the math says when you follow it.