This is one of the most useful heuristics in AI media literacy: you can predict the direction of a source's bias with reasonable accuracy just by knowing what kind of organization produces it. This doesn't mean those sources are lying — it means their incentives reliably shape what they cover, how they frame it, and what they leave out.

Investors and VCs need their portfolio companies and the overall sector to look promising. Their public communications are essentially long-form investment theses. When they write about AI capabilities, they're also, consciously or not, making the case for the sector they've bet on.

AI company communications — press releases, CEO posts, official blogs — are marketing materials with technical vocabulary. This doesn't make them useless. But the framing "our model achieves X on benchmark Y" requires you to know whether benchmark Y is actually meaningful, whether competitors were tested fairly, and whether the capability generalizes outside the test condition.

Tech journalists are incentivized by engagement, which rewards recency and strong takes. Their coverage is often the best available on breaking news but frequently lacks the sustained critical lens to evaluate whether the breaking development actually matters as much as reported.

Academic researchers have their own biases — toward their own area of focus, toward certain methodologies, sometimes toward prestige — but their work is subject to peer critique in ways that investor and corporate communications are not. The caveat: academic AI research is often months or years behind the commercial frontier.

Reliable doesn't mean neutral — it means that a source has demonstrated, over time, that it will tell you when it got something wrong, that it distinguishes between what is known and what is speculated, and that its past predictions have a track record you can actually check.

Some specific things to look for:

Explicit uncertainty language. Sources that use phrases like "the evidence suggests," "we don't yet know," and "earlier reporting on this was incorrect" are demonstrating calibration. Sources that consistently state things with high confidence regardless of how uncertain the underlying situation actually is are performing confidence, not reporting it.

Distinguishing between demos and deployment. A huge amount of AI hype lives in the gap between a demo (controlled, optimized, cherry-picked) and actual deployment (messy, variable, failure-prone). Reliable sources make this distinction explicitly. Unreliable ones show you the demo video and imply the deployment is equivalent.

Accuracy on previous claims. The single most useful signal for a source's reliability is whether their past confident claims turned out to be right. This takes effort to check but pays significant dividends. If a newsletter told you in 2023 that AGI was "eighteen months away," that's information about how much to trust their next prediction.

You don't need twelve newsletters. You probably need three to five sources that cover different angles: one that has good access to practitioners and tracks actual deployment (not just announcements), one that covers policy and regulatory developments, one that gives you academic-level depth on capabilities and limitations, and one that does critical journalism on AI industry claims.

The specific sources that fill these slots will change over time — publications get acquired, researchers shift focus, journalists change beats. What matters more than the specific list is that you're intentionally maintaining coverage diversity across these angles rather than accidentally following five sources that all share the same priors.

One practical exercise: take the sources you currently follow and categorize them by type. If they're all investor-adjacent or all skeptic-adjacent, you're getting a monoculture. A well-calibrated picture of AI requires input from people who have very different relationships to the technology and different stakes in its development.

There's a specific failure mode in AI coverage worth naming separately: access journalism. When a reporter gets exclusive interviews with AI company leadership, early access to new models, or briefings before public announcements, they gain reporting advantages — but they also take on implicit obligations. Biting the hand that feeds you access means losing the access.

This produces a structural dynamic in AI journalism: the reporters with the best access to the companies building AI tend to publish the most positive coverage of those companies. This isn't corruption in any simple sense — it's the predictable result of an incentive structure. Being aware of it means treating "exclusive access" pieces differently than independent analysis. The access piece is valuable for what the company wants you to know. The independent analysis is valuable for what the company doesn't.

You can't deep-dive every AI announcement. But you can apply a consistent triage process that takes less than five minutes and gives you a defensible initial position. Here's how it works:

Step 1: Who released it and what do they gain? A company releasing its own model and announcing it's revolutionary has obvious incentives. A third-party researcher testing that model and publishing results has different incentives. An academic lab replicating results has different incentives still. The announcement source alone tells you how much skepticism to bring.

Step 2: What's the specific capability claim, and what is it being compared to? "Better than GPT-4" tells you almost nothing without knowing what was measured, how it was measured, and whether the thing being measured is actually what matters for your use case. "Reduces hallucination rate by 60% on medical question-answering benchmarks" tells you something specific that you can reason about.

Step 3: Has anyone without a stake tried it yet? Early reports from people who have no affiliation with the company — even informal Twitter or Reddit posts from developers who clearly know what they're doing — are often more informative than official communications. Absence of these reports, especially for a widely-hyped release, is itself a data point.

Step 4: Does this change the inputs or outputs of something I actually do? Even if a claim is entirely accurate, it may not affect your work at all, or may not affect it yet. A new video model at $500/minute of generation is a real technology development, but if you're a freelance content creator it's not yet your problem. Relevance filtering is a legitimate part of triage, not avoidance.

Step 5: Set a review date, not an immediate verdict. For genuinely uncertain new developments, the best answer is often "I'll form a real opinion in three weeks when the initial hype cycle has played out and independent evaluations exist." This is not a cop-out — it's appropriate calibration. The pressure to have an instant take is real but not well-served by guessing.

The AI hype cycle has recurring patterns that are worth knowing because they let you calibrate your response before you have all the details. Here's the pattern that plays out most consistently:

Day 1–3 (announcement): Company announcement or leak. Extreme takes in both directions. "This changes everything" vs. "just another incremental improvement." Influencers and investors share enthusiastically. This is the lowest signal-to-noise moment. Anything you conclude here has high probability of needing revision.

Week 1–2 (initial testing): Developers, researchers, and power users who get access start sharing real results. This is where the gap between demo and deployment starts becoming visible. Capability limitations, failure cases, and real performance data begin emerging. This is when the picture starts to clarify.

Week 3–6 (consolidation): More systematic independent evaluation, academic commentary, and media retrospectives. By this point the actual capability delta — how much better is this really than what existed before — is usually assessable. This is the minimum point at which you should form a strong opinion about a new release.

Most people's AI information diet is heavily weighted toward Day 1–3 content, which is also the worst-calibrated. Deliberately waiting even a couple of weeks before forming a firm opinion will improve your accuracy substantially.

Sometimes you can't wait for the consolidation phase. If you're being asked to make a tool adoption decision at work, if a client needs immediate guidance, or if you're about to make an educational investment based on a capability claim, the stakes are real and the timeline is short.

In these cases, the five-minute triage protocol still applies, but you're prioritizing different signals: independent developer reports over official claims, conservative estimates over optimistic ones, and your own direct test of the tool over any documentation. If you can run a quick test yourself in a domain you understand well, that personal test is often worth more than ten reviews from people whose specific use case differs from yours.

One peer pattern worth flagging: a lot of people in our age range feel pressure to appear current and enthusiastic about new AI tools, which creates a bias toward early adoption and positive framing. There's social cost in being the person who says "let's wait and see." But "let's wait and see" is usually the right answer for major decisions, and building the confidence to say it is genuinely useful — professionally and personally.

Here's the long-game argument for building this habit: people who consistently apply good triage develop a track record. Over time, your colleagues, clients, and network notice whether your recommendations pan out. The person who excitedly endorsed three tools that didn't work as advertised loses credibility. The person who said "let's wait" and was later vindicated gains it.

Calibration compounds. Good judgment about one AI claim makes it easier to make good judgments about the next one because you start to recognize the patterns. You learn which types of claims from which types of sources have historically been accurate, and you build genuine expertise not about AI per se but about how AI claims work and fail. That expertise is durable in a way that specific tool knowledge isn't.

We've talked a lot in this module about staying current. But there's a different and more important question: what do you want to have built, in terms of knowledge and judgment, ten years from now? Because what you build over the next decade isn't just a set of current skills — it's a way of thinking that will shape how you navigate whatever comes after the current moment.

Here's what compounds over time and stays useful regardless of what specific AI technologies exist:

A reliable mental model of incentive structures. Understanding who benefits from AI hype, who has incentives to understate capabilities, and how those incentives shape public discourse — this knowledge doesn't expire. The specific actors change, but the incentive map stays structurally similar.

A track record of calibrated predictions. Every time you make an explicit prediction about an AI development and then check whether you were right, you're building a personal calibration database. Over years, this becomes a genuine signal about your own cognitive biases and blind spots. It's uncomfortable in the short term and extremely valuable in the long term.

Comfort with deep uncertainty. AI development involves real epistemic uncertainty — we don't know how capable these systems will become, we don't fully understand how they work, and experts disagree profoundly about near-future trajectories. People who can hold genuine uncertainty without either collapsing into nihilism or papering it over with false confidence are more useful in this domain than people who can't.

A "personal AI position" is a set of considered views about how AI is likely to develop and what that means for your specific life — your career field, your creative work, your financial decisions. This is different from following the news. It's an active synthesis that you update deliberately as new evidence comes in.

A personal position has a few components. It includes your best current assessment of AI capabilities in your specific domain — not in general, but in the specific applications that are relevant to you. It includes your view on the timeline for those capabilities to mature or change. And it includes explicit acknowledgment of what you're most uncertain about and what evidence would change your view.

The reason to build this explicitly rather than leaving it as a vague intuition is that explicit positions are updatable. When you've written down "I think AI writing assistance will be table stakes for marketing roles within two years," you can check that belief against reality, notice when evidence undermines it, and update it. Vague intuitions don't update cleanly — they shift imperceptibly and often in ways you don't notice.

Your peers who are doing this well — and some are — tend to be the ones who are most useful to talk to about AI. They don't just share hot takes; they share reasoning. They know what would change their mind. Those qualities are rarer than they should be and significantly more valuable professionally than simply being enthusiastic about new tools.

There's a specific cognitive habit that separates people who get better at evaluating AI over time from people who just accumulate more opinions: the practice of explicitly updating beliefs when evidence warrants it, and being willing to say out loud "I was wrong about this."

This is harder than it sounds. The social dynamics around AI are intense — there are strong tribal affiliations between AI skeptics and AI boosters, and both sides treat changing your position as evidence of weakness rather than evidence of learning. This is exactly backwards. In a genuinely uncertain and fast-moving domain, the people who update most readily when evidence changes are the ones building the most accurate picture over time.

One concrete practice: keep a short log of your AI-related beliefs and predictions. Once a month, review the log and note where reality has diverged from what you expected. Don't be harsh on yourself about being wrong — being wrong in interesting ways is how you learn the actual patterns. Be curious about why you were wrong. What did you misread about the incentives? What did you not know about how the technology actually works? These post-mortems are more useful than any amount of general AI reading.

Everything in this module — the information half-life concept, the source stack, the triage protocol, the personal position — is in service of a single thing: developing the kind of relationship with AI that puts you in a position of active judgment rather than passive reaction.

You're going to spend your career in an environment shaped significantly by AI. That's not a prediction anymore — it's already the case. The question is whether you navigate that environment with a reliable compass or just respond to whatever the loudest signal happens to be at any given moment.

The people who will do best in this environment aren't necessarily the ones who know the most about AI right now. They're the ones who have built the habits that keep their knowledge accurate and their judgment well-calibrated over years. That's what this module has been about. Not a snapshot of the current state — a practice for staying clear-eyed in a domain that will keep moving.

You've put in the work. Now the real project is maintaining it.