The Delulu Blog

AI Shouldn't Decide What's True—It Should Show You What You're Missing

By Eric Barker · August 2026 AI & Automation

When you encounter a news story today, the information flow looks roughly like this: event happens → journalist reports it → article gets published → you read it and decide what you think.

The reader is responsible for supplying everything else: the historical context the article didn't include, the competing interpretations that might fit the same evidence, the assumptions embedded in the conclusion, what actually happened the last time this argument was made, and whether a prediction about tomorrow is being improperly extrapolated from a fact about today. That's a lot to hold in your head while you're also working, raising kids, and keeping up with everything else that demands your attention.

A sufficiently capable personal AI could eventually provide that missing layer automatically.

An example: the AI data center problem that might not be permanent

Take a concrete case. Articles about AI data centers and electricity consumption are accurate about the present. Newer facilities do consume enormous amounts of power and water. Grid pressure is real. Those facts shouldn't be dismissed or softened.

But reporting sometimes leaves an implicit assumption embedded so deeply that readers absorb it without noticing it's even there: today's infrastructure requirements equal AI's permanent infrastructure requirements. That assumption sits underneath much of the coverage, and it may or may not prove correct.

What if a good context layer could intercede? Not by saying "this article is propaganda"—it's not. Instead, something like:

The power and water figures you're reading are probably accurate measurements of current consumption. However, this article projects them forward without separating current fact from future assumption. Local inference models, quantization, dedicated hardware accelerators on-device, and energy-per-inference improvements are all advancing rapidly. The article you're reading assumes centralized inference remains the default. Here's the evidence supporting that assumption. Here's the evidence suggesting it might shift. Here's what would have to happen for each scenario to occur. Here's what researchers disagree about. Here's what remains unknown.

Notice what this does and doesn't do. It doesn't label the original article as false or propaganda. It doesn't declare centralized data centers doomed. It doesn't inject ideology. It makes the argument legible—it excavates the assumptions lying underneath the conclusion and puts them where the reader can actually see and evaluate them.

That's a very different operation from what most "AI fact-checking" claims to do.

What we usually mean by "AI fact-checking" (and why it's backwards)

The standard pitch for AI fact-checking goes roughly: AI will evaluate claims and label them TRUE, FALSE, or MISLEADING. It's appealing because it's simple and looks decisive. It also doesn't work and can actively make things worse.

A system that automatically labels information with high confidence often just replaces human gatekeepers with algorithmic ones. And algorithmic gatekeepers have failure modes humans at least understand how to spot: training-data bias, source-selection bias, model provider influence, politically-tilted training, overconfidence in uncertain territory.

More fundamentally, it misunderstands the shape of real arguments. Most important claims don't reduce to binary verdicts. They're embedded in assumptions. They make implicit predictions. They depend on which timescale you're asking about. They rest on who gets to decide what counts as evidence. Whether AI's infrastructure footprint matters for the grid depends on whether you're asking about 2026 (current measurements matter) or 2031 (technological change matters). Both timescales are reasonable questions. A TRUE/FALSE verdict collapses that distinction into nonsense.

The better architecture: context, not truth

Imagine instead a system that makes arguments more legible by exposing what usually stays hidden:

The assumptions. What must be true for this conclusion to follow? If an article claims technology X will displace a million jobs, is it measuring permanent-job-loss or task-displacement? Is it assuming no new job creation? Over what timescale? Those assumptions aren't incidental—they're foundational to whether the claim makes sense at all.

The strongest evidence supporting it. Not a cherry-picked example, but the single best argument someone who believes this could make. What's the honest case?

The strongest credible counterargument. Not a weak man, but a genuine steelman: the best case someone skeptical could make. What would a thoughtful person who disagrees actually say?

What was left out. What relevant information didn't make it into the article? What context would complicate the story?

The incentives. Who benefits from this interpretation? Does that invalidate the evidence? No—but the reader should know who's paying for the research and whether the researcher has skin in the game.

Uncertainty. What don't we actually know? Where do credible experts disagree?

Historical parallel. Has this argument been made before? What happened the last time? Did the prediction come true? Where was it wrong?

Falsifiability. What would it take to prove this prediction wrong? What future evidence would settle the question?

Notice what a system doing this is not doing: it's not deciding what's true. It's making the argument transparent so a human can actually decide what they believe, equipped with better information.

The most important safeguard: challenge the reader, too

A personal AI could become the most sophisticated confirmation-bias machine humanity has ever created if it simply learns what you believe and generates endless supporting evidence.

The system has to be capable of pushing back on the reader's own assumptions with the same rigor it applies to the article. If you encounter an article claiming AI data centers prove the technology is unsustainable, and you respond "obviously true, this proves the critics were right all along," a good system should say: wait. Current measurements can be legitimate without validating every extrapolation. Data centers do use real electricity and water. But local inference is also accelerating. And we don't know yet whether centralized or distributed deployment will prove more efficient. You're oversimplifying this as much as the tech-optimist who dismisses environmental concerns entirely.

That matters because your own assumptions are often the hardest ones to examine.

The game-changer: persistent memory

This is where the architecture gets genuinely different from existing tools.

Today's fact-checking largely works on individual claims in isolation. A system evaluates one statement and renders a verdict, separate from everything else.

But imagine a personal AI that remembers. You encounter an article in August 2024 claiming that local AI models will never match frontier capability. Your AI stores not just the claim but the prediction embedded in it, the evidence available at the time, and your own skepticism or agreement.

Then in August 2026, you encounter another article on the same topic. Your AI says: two years ago, you read a similar prediction. Here's what has changed since then. That forecast was partially right and partially wrong. Here's what your earlier skepticism missed. Here's where that article got it exactly right. Here's what happened with the three-year timeline people predicted.

This creates something conventional fact-checking can never do economically at scale: longitudinal accountability. Claims aren't evaluated once in isolation. They're evaluated over time, against reality, with a historical record of what people believed when.

That transforms the conversation. If someone makes a prediction about local AI capability in 2026, you can revisit that prediction in 2027 with real evidence. If journalists make assumptions about infrastructure in 2024, you encounter those same assumptions in 2028 with years of follow-up data to compare them against.

That's not possible with centralized fact-checking services. But it's exactly what a personal system with persistent memory could do.

Why this matters across domains

The examples so far have centered on AI and technology. But the architecture applies everywhere arguments appear:

Economics. A headline claims interest-rate policy will cause a recession. Your AI surfaces: what timeframe? What does "recession" mean here? Who benefits if this prediction comes true? What happened when policymakers raised rates at this particular inflation level before?

Climate and environmental. An article predicts sea-level rise scenarios. Your AI excavates: these are model predictions, not measurements. Here's the range of credible expert disagreement. Here's what's actually measured vs. what's projected. Here's what would falsify this model.

Product claims. A company announces a breakthrough technology. Your AI surfaces: who funded this research? Do they have competing interests? What's the real performance claim vs. the marketing framing? What assumptions does "breakthrough" rest on?

Policy. A politician claims a policy will create X jobs. Your AI asks: temporary or permanent? Counting what as employment? Based on what model? What assumptions about complementary changes?

The common thread: most arguments leave critical reasoning work to the reader, usually without the reader realizing it. A good context layer makes that work visible and available.

The failure modes are real and worth taking seriously

This entire framing assumes the system is honest and capable. It isn't, not yet, and several failure modes deserve their own serious attention:

Hallucination. Models make up citations that sound plausible. A context layer that confidently presents fabricated sources is worse than no context at all.

Bias in training data. If the model was trained on text that systematically favors certain interpretations, the "counterargument" it surfaces might be a strawman, not a steelman.

Filter bubbles from personalization. A system that learns what you believe might gradually learn to show you only interpretations that align with your existing views, making the context layer a confirmation-bias amplifier.

Source selection bias. Which evidence does the system consider "credible"? Who decided? Does that decision systematically exclude viewpoints?

Model provider influence. Who built this system? Do they have incentives that shape what "context" looks like?

Overreliance on AI. If people defer to the context layer instead of thinking for themselves, it becomes a form of automation bias where AI reasoning substitutes for human judgment.

Privacy and surveillance. A system with this much visibility into what you read, what you believe, and how you change your mind is an unprecedented surveillance tool if it's centralized.

These aren't edge cases. They're foundational architectural problems. Which brings us to the prerequisite.

Why this requires local models to work

As AI capability increasingly lives on personal devices rather than centralized services, a truly trustworthy personal epistemic assistant becomes possible. And becomes necessary.

A context layer that works at this level of intimacy—storing what you've read, what you believed, what you predicted, where you changed your mind—is a continuous record of your intellectual history. Centralizing that data creates obvious risks: surveillance, manipulation, breach exposure, government request leverage, and the simple fact that whoever owns the server owns your thoughts.

An on-device system avoids those risks by architectural design. Your intellectual history stays yours. The reasoning stays auditable locally. The filter-bubble risk drops because you control the training signal.

This isn't a theoretical advantage—it shapes whether the system is trustworthy at all. A centralized epistemic assistant is a contradiction. A personal one becomes possible once the models are capable enough to run locally, which we're rapidly approaching for this use case.