AI Has Endless Answers But A Shrinking Supply Of Human Truth
Progress in the AI realm has historically focused on measuring chip efficiency and monitoring parameters; however, as the internet that trained today’s models has become full of text written by those same models, a new constraint is becoming increasingly visible: the supply of original, verifiable human data.
SEO firm Graphite, for instance, recently went through over fifty thousand English-language articles from Common Crawl and found that roughly half of them were AI-generated. That’s not all; that trend (of AI content flooding the interweb) has been prevalent since the beginning of last year and shows no signs of slowing down any time soon.
On the subject, Trevor Koverko, co-founder of Sapien, a platform that pays people to label and verify AI training data, believes that as the internet continues to saturate with AI-generated data, “a fresh observation from a real person becomes more valuable,” adding:
“Models can generate endless plausible answers. They still need evidence of what actually happened, what worked and where they got it wrong.”
When Models Learn From Themselves
The risk of leaning on machine-made data is well documented , and in a joint research paper published by Oxford and Cambridge, it was made clear that the indiscriminate use of model-generated content in training is causing “irreversible defects in resultant models.”
Not only that, but the team also concluded that data from genuine human interactions is becoming increasingly rare and valuable, especially with the prevalence of LLM-generated content continuing to seep into what gets crawled from the web.
Koverko is wary of diagnosing collapse in a commercial model from the outside, but the symptoms, in his view, are familiar, i.e., polished, repetitive answers that miss unusual cases. Synthetic data works where answers can be checked against code, math, or a simulator, and far less well where they depend on judgment.
The Trust Gap Shows Up in the Survey Data
That technical problem has a consumer-facing twin because just last month, according to a recent survey from my company, Prosper Insights & Analytics , nearly 40% of all adults in the US actively make use of generative AI (for a myriad of purposes ranging from research work to simple online searches).
And it’s not like users are not aware of their dependency; the Prosper Insights & Analytics survey showed that over one-third of adults believe that AI needs human oversight, while an equally large population believes that AI routinely provides wrong information. Other worries include:
- Disclosure about the data AI uses
- Lack of proper regulation
- Trust deficit, especially when it comes to AI models and their creators having the best interests of users in mind
- Extreme concern regarding AI violating individual privacy
The generational splits are extremely stark, as well: among Boomers, 52.1% say AI needs human oversight, 47.9% worry about hallucinations, and 42.5% don’t trust it to act in their interest. Millennials are the least worried on all three counts.
Even then, roughly one in three Gen-Zers have flagged both oversight (32.7%) and wrong answers (33.4%) as concerns for them as the space continues to mature and grow.
In Koverko’s reading, the numbers are a demand for recourse rather than reassurance, adding:
“Those survey numbers tell me people want a way to challenge an answer, not just a promise that a human was somewhere in the training process. Human oversight only earns trust when it produces fewer errors, clear accountability, and a route to correct a bad result.”
Handing an Agent the Wallet
Money is seemingly where hesitation among users peaks, as Prosper found that just 7.2% of adults would use an AI agent to invest in stocks, bonds, or crypto, and 8.2% would let one pay their bills.
Moreover, nearly six in ten users would not use agentic AI for anything, a figure that climbs to 81.7% among Boomers.
Generation matters here too, as 27.9% of Millennials think agentic AI is a good idea, compared with just 4.9% of Boomers.
Lastly, only 10.8% of Millennials would let an agent invest for them (against 2.4% of Boomers).
Katerina Vdovichenko, co-founder of ORO, which builds AI tools for managing digital assets, doesn’t see the reluctance as technophobia. Instead, she pointed out:
“Those numbers aren’t a failure of imagination. They’re the public doing the risk assessment our industry skipped. People don’t doubt that AI is smart. They doubt that anyone has defined what the agent is allowed to do.”
Furthermore, she doesn’t think the generational gap is really about age, or that it will close on its own. In her opinion, the boundary for autonomy should be set by authority rather than size, adding that “most people draw the line at dollar amounts, and that’s the wrong place.”
Routine actions inside a mandate the user has already set can become automatic over time, but anything irreversible, and anything that expands the agent’s authority, should always need a human, in Vdovichenko’s outlook.
Lastly, transparency means clarity rather than volume, and before an agent acts, a user should see four things in plain language:
- What it intends to do
- What it will cost
- What the worst case is
- What can’t be undone.
Afterward comes a receipt the agent didn’t write itself, because nobody trusts a company that audits its own books. Onchain finance offers one version of that receipt, since transactions settle on public ledgers anyone can inspect, though Vdovichenko is careful not to oversell it, adding:
“A blockchain is a perfect witness and a terrible judge. It proves what happened, but not why, and not whether the action was authorized.”
Working Toward a Holistic AI Future
The same fluency problem runs back into training too, and its sharpest residuals arise in specialized fields. To this point, Stanford and Yale researchers tested AI legal research tools from LexisNexis and Thomson Reuters and found that they each hallucinated more than 17% of the time.
In sum, models need human data that can be traced and checked, and consumers want proof of who checked the work and a way to object when it goes wrong. For executives deploying AI, Koverko suggests one question for vendors:
“Show me how you know this system is right on the cases that matter to my business, who checked those cases, and what happens when it gets one wrong.”
A vendor that can only show a broad benchmark score, he added, has not answered it. Vdovichenko expects the same test to decide which financial AI products win. “In five years, every financial agent will be smart,” she said. “The trusted ones will be the ones that knew where to stop.”
Disclosure: The consumer sentiment study referenced above was conducted by my company, Prosper Insights & Analytics . This is the same dataset used by the National Retail Federation, and available from Amazon Web Services, Databricks, and the London Stock Exchange Group for economic benchmarking.