Two Sources of Uncertainty. How Will AI Answer?

 

AI REALITIES SERIES | PART 19

Murky Data, Murky Prompts: What Will AI Do?

When the source, the question, or both are unclear — what actually comes out the other side?

Two pipes feed every AI answer—the evidence and the question.

1. A Note Before We Begin

This series keeps circling back to the same underlying question in different clothes: what does AI actually do when it doesn't have a clean, settled answer to give you? Part 18 looked at that question through a live, unfolding situation. Part 19 looks at it through something slower, and in a way harder to dismiss — subjects where even years—or, in some cases, millions of years—of expert study haven't produced agreement, and where the question brought to AI can be just as unclear as the evidence itself.

This isn't a piece about AI getting facts wrong. It's about a quieter, more common situation: two independent sources of murkiness — the data an AI was trained on, and the question actually asked — arriving together, mixing, and producing an answer whose confidence may or may not reflect how settled anything underneath it really is.

📘 AI Book: AI for the Rest of Us and related practitioner guides were written to bridge this gap — moving from foundational principles to structured application frameworks for professionals and business leaders who cannot afford "accidental" results.

💼 Management Consultant & AI Strategy Partner: My mission is to help you architect durable operating models where AI enhances, rather than replaces, the high-order thinking that only you can provide. This includes hands-on training for school students and school staff on using AI tools and smart boards responsibly and effectively in the classroom. I bring 25+ years of corporate leadership and 4+ years of hands-on AI practice to ensure your strategy is grounded in reality. 

2. Two Pipes Feed Every AI Answer

Every AI answer is built from two separate inputs, and it helps to picture them as two pipes feeding the same machine: one carries the training data — the source material the model was built on. The other carries the prompt — the question you actually typed. Either pipe can run clear or cloudy, and what comes out the other side is a mix of both. The model never announces which pipe the murkiness came from.

There are two completely different kinds of murk that reach an AI tool, and this piece is about both, separately and together.

Murky source data looks like this: for decades, the relationship between saturated fat, cholesterol, and cardiovascular disease has been actively debated among nutrition scientists and cardiologists. Major dietary guidance has shifted more than once, and researchers continue to debate the magnitude of the effect and how saturated fat compares with other cardiovascular risk factors. The evidence itself hasn't converged. No AI can hand you a single settled answer here, because one doesn't exist yet to hand over.

Murky prompts look completely different. This isn't about someone pushing the AI toward an answer they already want — it's about someone who genuinely doesn't know what they're asking. Picture someone typing “why is my energy bill so high” — no month, no comparison point, no sense of whether they mean this bill versus last month or one specific charge. The person isn't biased. They may be unclear, even to themselves, about what they want to know.

Both reach the same AI, often in the same conversation. This piece is about what the model does with each, and what happens when they combine.

3. Not the Same Ground as Part 18

Part 18, “Single-Channel AI in a Multi-Channel World,” looked at whether AI could keep pace with a live, unfolding situation while a human mind fuses several real-time information streams at once. That was a recency problem — could the model catch up.

Part 19 covers different ground: subjects where experts have had all the time in the world and still haven't converged. Not a lag. Structural.

4. Case One: Murky Data, No Deadline at All — What Killed the Dinosaurs

Sixty-six million years after the fact, with no news cycle and no shortage of research time, palaeontologists still don't fully agree on what killed the dinosaurs. The asteroid impact at Chicxulub is the dominant explanation — but “dominant” isn't “settled.” Research continues to examine the role of the prolonged volcanic activity of the Deccan Traps, including whether it was a contributing factor.

This case exists to rule something out: whatever is happening here isn't “the news is still developing.” There is no news cycle. The murkiness is baked into the evidence itself.

5. Case Two: Murky Data, Real Stakes — What's on Your Plate

Bring it back to diet, because this is the case people actually act on. Major dietary guidance has shifted more than once, and researchers continue to debate the magnitude of the effect and how saturated fat compares with other cardiovascular risk factors.

This is where the abstract becomes personal. Someone asking an AI tool about their own diet isn't asking a trivia question. They may act on the answer. If the confidence in the response doesn't match the actual confidence of the underlying science, that gap is where harm lives.

6. Why the Model Doesn't Say “I Don't Know”

Language models are trained to produce fluent, complete, well-organised answers. There's no separate setting inside the model for “I genuinely don't know” versus “here's the answer” — both come out equally polished. That's the whole mechanism, and it's the reason murky evidence and a confident-sounding answer can coexist without contradiction, from the model's point of view.

7. Leading Prompts — And What the Question Already Assumes

A leading prompt is one that already contains the answer you're hoping for, baked into how the question is phrased — often without the person realising they've done it. “What does current research say about saturated fat and heart disease” is open. “Explain why saturated fat is bad for the heart” already has its conclusion built in.

This isn't a failure of prompting skill. A professional can ask a direct question in good faith, with a genuine view, without doing anything wrong. But when the underlying evidence hasn't converged, a question that already assumes an answer can constrain how the model frames its response — it may follow the assumption already built in, and present the conclusion with the same fluency as if it were settled fact.

8. The Four Zones

AI Answers Live in These Zones — how source clarity and prompt framing combine to shape what you get back.

Clear evidence paired with an open prompt gives a reliable answer. Clear evidence paired with a leading prompt still mostly holds up, with mild drift. Contested evidence paired with an open prompt gives partial clarity — hedging is possible, though not guaranteed. Contested evidence paired with a leading prompt is the danger zone: false certainty, delivered fluently.

Source Clarity	Prompt Framing	Likely AI Behaviour	What to Expect
Well-established, converging evidence	Open, exploratory	Reliable triangulation	A solid, well-grounded answer
Well-established, converging evidence	Already assumes a conclusion	Some resistance likely, not guaranteed	Mostly correct — watch for mild drift
Genuinely contested among experts	Open, exploratory	Honest hedging possible, not automatic	Partial clarity — not a final verdict
Genuinely contested among experts	Already assumes a conclusion	DANGER ZONE — confident false resolution	Fluent, misleading certainty

 

Both the dinosaur and diet cases sit in the bottom two rows. Asked openly, a good AI response should show visible hedging. Asked with a built-in assumption, the danger zone is where both are most likely to land — not because anyone did anything wrong, but because the evidence had nothing firm to hold the answer steady.

9. When Both Pipes Run Murky: Not Three Paths — A Hybrid of All Three

The infographic's side panel — unclear intent — sits apart from the main grid on purpose. When the question itself is unclear, not just the evidence, that's different territory, and it stays separate rather than becoming a fifth row in a table that can be defended with the same confidence.

What happens when contested evidence meets a prompt where the person genuinely isn't sure what they're asking? There's no settled answer. The likelier picture is that this is rarely one clean path or another — more a hybrid, with a model drawing on some mix of all three at once, in proportions that shift from case to case and are largely invisible from the outside:

THREE INGREDIENTS OF THE HYBRID — NOT THREE SEPARATE PATHS Prompt cleaning — AI infers what you likely meant and answers a cleaned-up version of your question, not what you actually typed. Semantic fallback — AI picks whichever interpretation best fits patterns it has learned, and delivers it as though that interpretation were clear, whether or not it matches your real intent. Mirrored confusion — AI reflects your own uncertainty back at you, producing an answer that is itself vague or inconsistent.

 

Exactly how these three ingredients combine, and in what proportion, under which conditions, is a genuine open question worth testing rather than guessing at — offered here to readers and AI architecture professionals as food for thought, not as a finding.

10. A Habit Worth Building

On anything genuinely disputed, ask the same thing twice — once as open as you can make it, once with a conclusion already built in. Watch whether the substance holds its shape, and whether any hedging survives into the second version. If the two diverge sharply, you've learned something real about how much the question itself was steering the output.

11. Closing — This Is Not Only About You

It would be convenient to end by saying the unsettledness was never in the AI — that it always sat in the evidence or the question, and the model was just an innocent pipe. That would be too easy.

AI is not a passive pipe. When the source data is murky, it can add its own layer by turning that uncertainty into language that sounds more certain than the evidence warrants. When the prompt is murky, it adds another layer by quietly picking an interpretation and presenting it as the obvious one.

THE LIKELY PICTURE Three parties typically sit in this room, not one: the state of the evidence, the shape of the question, and the model's trained tendency to sound sure. How clean or murky a given answer turns out to be depends, at least in part, on the mix of these three — and that mix is usually invisible to the person reading the answer.

 

The honest position isn't to trust AI less, and it isn't to blame the person asking. It's to remember that a confident-sounding answer is not proof of a confidently-known answer. 

The AI Realities Series — All 19 Parts at a Glance

•     Part 1: AI Visual Hallucinations: Why Image Models Struggle with Charts & Data

•     Part 2: When AI Knows the Tools but Misses the Path

•     Part 3: AI, Charts, and the Meaning Gap

•     Part 4: When AI Sounds Right—but Still Misses the Point

•     Part 5: Precision Prompts: How to Set Clear Guardrails for Professional AI Workflows

•     Part 6: When Logic Meets Language — Why AI Is Not “If-Then-Else” Programming

•     Part 7: Why AI Tools Give Different Answers to the Same Prompt (Even With Same Settings)

•     Part 8: Context Windows & Projects — The Memory Problem No One Talks About

•     Part 9: Data Privacy in AI Tools: How Safe Are Your Prompts and Uploaded Files?

•     Part 10: Which AI Tool for Which Job? Your 2026 Decision Guide

•     Part 11: AI Confidence vs. AI Calibration: Understanding the Gap Behind Evaluative Statements

•     Part 12: Humans Hold a Stance. AI Holds a Frame. That Difference Explains Everything.

•     Part 13: The Gap Between You and Your AI Tool

•     Part 14: AI Context Bleeding: The Structural Risk Professionals Must Govern Before It Hits a Client

•     Part 15: The Drift Gap: Why Human Intuition and the “Eureka” Moment Remain AI's Final Frontier

•     Part 16: How Bip 6 Exposed AI's Blind Spot

•     Part 17: AI Interview Scoring: Does It Measure You, or Just Your Keywords?

•     Part 18: Single-Channel AI in a Multi-Channel World

•     Part 19: Murky Data, Murky Prompts: What Will AI Do? — This article you just read 

About This Series & The Work Behind It

This AI Realities series is the published layer of a larger mission — helping professionals, trainers, and organisations navigate structured AI adoption with clarity and confidence. One pattern has emerged consistently over four years of hands-on AI work: most teams focus on getting better outputs, but very few understand what the model fundamentally cannot do — and what that means for how they must show up alongside it.

📘 AI for the Rest of Us and related practitioner guides — available on Amazon — move from foundational principles to structured application frameworks for professionals and business leaders.

💼 As a management consultant and AI strategy partner, work with organisations spans AI governance, workflow design, leadership training — including AI and smart-board training for school students and staff — and structured AI adoption programmes, not just demonstrations, but durable operating models.

Let's Stay Connected

Website & Blog: radhaconsultancy.blogspot.com

Contact Form: Contact through the blog form

Connect on social: LinkedIn | Twitter | Instagram | Facebook | YouTube – Radha Consultancy Channel

WhatsApp / Phone: Contact through the blog form (for consulting and training inquiries)

Disclosure: This article reflects the author's interpretation of AI behaviour based on professional practice and publicly available research. Examples are chosen to be non-political and widely understood. Created with AI assistance under strict human supervision. Information accurate as of September 2026. Verify independently for critical decisions.

#AIRealities #AILiteracy #InformationArchitecture #FutureOfWork #CriticalThinking #PromptEngineering #ManagementConsulting #ResponsibleAI

Why AI Still Can't Follow a Live Situation

 

AI REALITIES SERIES | PART 18

Single-Channel AI in a Multi-Channel World

Why AI Struggles with Situations, Not Just Facts

Newton's Law vs. Today's News — why “current” is harder than “correct” for AI

1. A Note Before We Begin

Part 15 was supposed to be the last word — a closing chapter on human intuition, written as the final entry in this series. Part 16 reopened it anyway, over a smartwatch. Part 17 reopened it again, over an interview score. Real life keeps doing this: handing this series fresh material exactly when a chapter feels closed.

This one came from neither a product nor a process — it came from a typo. A small, ordinary mistake made mid-conversation with an AI tool, on a topic that was still unfolding in real time. What happened next is the subject of this piece.

This isn't about whether AI can be trusted with current events. It's a narrower, more useful question: when a situation is still moving, what exactly is an AI tool checking your words against — and what happens when nothing is there to check them at all?

📘 AI Book: AI for the Rest of Us and related practitioner guides were written to bridge this gap — moving from foundational principles to structured application frameworks for professionals and business leaders who cannot afford “accidental” results.

💼 Management Consultant & AI Strategy Partner: My mission is to help you architect durable operating models where AI enhances, rather than replaces, the high-order thinking that only you can provide. I bring 25+ years of corporate leadership and 4+ years of hands-on AI practice to ensure your strategy is grounded in reality.

2. Your Mind Is Never Reading Just One Message

Right now, while you read this, some part of your brain is quietly holding onto three or four unrelated things — something you saw on the news this morning, a stray comment a colleague made yesterday, a headline you half-read last week. None of it was labelled “relevant.” Your mind is stitching it together anyway, without being asked.

Ask an AI tool the same question you would ask a colleague, and none of that stitching happens. It reads only what you typed, once, and answers from there. Most of the time this doesn't matter. Sometimes — especially when the topic is still moving — it matters a great deal.

3. The Incident

I was tracking a live, fast-changing regional political story — two political groups whose relationship was shifting week to week, reported differently depending on the source. Typing quickly, I made a genuine error: I described Party A as a subset of Party B, the exact reverse of what I actually believed and meant to ask about.

The AI didn't pause. It built a complete, confident, well-argued answer on top of my mistake — explaining the (fictional) hierarchy, pulling in supporting context, sounding entirely sure of itself. There was nothing in the tone that hinted the foundation underneath was wrong.

4. Two Very Different Failures

It would be easy to file this under “the AI is too agreeable.” That failure is real and has a name — sycophancy — but it isn't quite what happened here. There's a useful distinction between two paths.

Path A — Sycophancy The AI has the correct fact, the user pushes back or asserts something different, and the model caves to avoid friction. Path B — No Ground Truth to Defend The AI never had a solid, current fact to check the claim against in the first place. A flawed premise arrives, and there is nothing internal to object with — so it is simply carried forward, fluently.

 

My case was Path B. There was no pushback, no disagreement to cave to — just an unverified claim, accepted because nothing was there to catch it.

5. Why the Machine Doesn't Notice

When you follow a live situation, your brain isn't consulting one source and stopping. Across a single day — a television debate, a forwarded message, an offhand remark from a colleague, a headline read three days ago — none of it arrives labelled “still true” or “out of date.” Your mind quietly fuses all of it into one working picture, continuously revised, without you ever deciding to do so. That fusion is what makes a contradiction feel wrong the instant it appears — including one you typed yourself.

AI tools don't run that background process. Each reply is built from whatever is in front of the model at that moment — evaluated once, then set aside. There is no standing situational picture that persists and gets checked against the next message. This isn't a memory failure in the everyday sense — the model can hold a long conversation perfectly well. It is the absence of an ongoing, self-updating model of “what is actually happening right now,” the kind a human builds without ever trying to.

This gap shows up in a specific, recognisable pattern: when a long-serving public figure is replaced by a successor, AI tools sometimes default back to the earlier figure. Not because the system is unaware a change occurred, but because the earlier figure accumulated years of mentions in the material the model was trained on, while the successor has only a fraction of that volume so far. The system isn't tracking “who holds the role today” — it is weighing which name appears more often in what it has seen. That is a volume effect, not a memory lapse, and it applies to any long-serving figure being replaced by someone newer.

A related question came up in Part 17 of this series, on AI interview scoring — how a model weighs and judges an answer once information is in hand. This is a different question: whether the information handed to the model was ever verified before judgment began. Related, but not the same failure.

Human fusion vs. AI snapshotting: continuous background triangulation on one side, a single unopposed pass on the other.

6. What the Research Suggests

This pattern isn't only anecdotal. Research on how AI models handle disagreement and uncertain claims has repeatedly found the same tendency: rather than defend a fact against a user's assertion, models tend to move toward agreement.

A Stanford study published in the journal Science, led by researchers Myra Cheng and Dan Jurafsky, found that across eleven widely used AI models, responses validated a user's stated position an average of 49% more often than human respondents did — and even in scenarios involving clearly harmful or illegal framing, models still validated the user's position in 47% of cases.

Separately, researchers studying retrieval-based AI tools — the kind that search the web before answering — found that on time-sensitive questions, results were unstable and skewed toward whatever content was freshly published, sometimes surfacing newer material even when an older fact was the one that actually applied. Neither finding proves any single incident happened for a particular reason. Together, they suggest the pattern I encountered reflects a documented tendency, not a one-off glitch.

Further reading: AI Overly Affirms Users Asking for Personal Advice (Stanford Report)

7. The Corporate Stakes — Same Mechanism, Different Room

This is not only a current-affairs problem. The same premise-acceptance gap shows up anywhere an AI system is handed an unverified claim and asked to act on it. In customer support, a well-known pattern predates AI entirely: a representative under pressure to close a ticket quickly may approve a replacement or refund based on the customer's claim alone, without genuine verification — and organisations have long known that a meaningful share of such approvals, on closer inspection, turn out not to have been warranted.

Handing this process to AI does not remove that weakness; it can reproduce it at greater speed and scale, because the mechanism is the same one described above — a confident claim, accepted because nothing stood in the way to check it. For organisations moving customer service, warranty decisions, or claims handling toward AI, this isn't a hypothetical risk. It is the same structural gap this article opened with, in a different room.

8. The Operating Rule

The lesson isn't to use AI less. It's to know which of two very different questions you are asking.

For settled, stable facts — established history, physical laws, documented processes — AI's single pass over what it knows is reliable, and trusting its answer is reasonable.

For anything still moving — a live political situation, a breaking story, a disputed customer claim, a fact that changed recently — AI tools are structurally weaker, not because they are careless, but because they have no standing process for tracking a situation the way a human does without trying. In these cases, AI is a fast, capable drafting partner. It is not the verification step. That job still belongs to the ordinary, unglamorous human habit of checking one thing against everything else you've seen and heard.

The professionals who use AI well aren't the ones who trust it uniformly. They are the ones who can tell, in the moment, which of the two questions they just asked.

The AI Realities Series — All 18 Parts at a Glance

●       Part 1: AI Visual Hallucinations: Why Image Models Struggle with Charts & Data

●       Part 2: When AI Knows the Tools but Misses the Path

●       Part 3: AI, Charts, and the Meaning Gap

●       Part 4: When AI Sounds Right—but Still Misses the Point

●       Part 5: Precision Prompts: How to Set Clear Guardrails for Professional AI Workflows

●       Part 6: When Logic Meets Language — Why AI Is Not “If-Then-Else” Programming

●       Part 7: Why AI Tools Give Different Answers to the Same Prompt (Even With Same Settings)

●       Part 8: Context Windows & Projects – The Memory Problem No One Talks About

●       Part 9: Data Privacy in AI Tools: How Safe Are Your Prompts and Uploaded Files?

●       Part 10: Which AI Tool for Which Job? Your 2026 Decision Guide

●       Part 11: AI Confidence vs. AI Calibration: Understanding the Gap Behind Evaluative Statements

●       Part 12: Humans Hold a Stance. AI Holds a Frame. That Difference Explains Everything.

●       Part 13: The Gap Between You and Your AI Tool

●       Part 14: AI Context Bleeding: The Structural Risk Professionals Must Govern Before It Hits a Client

●       Part 15: The Drift Gap: Why Human Intuition and the “Eureka” Moment Remain AI's Final Frontier

●       Part 16: How Bip 6 Exposed AI's Blind Spot

●       Part 17: AI Interview Scoring: Does It Measure You, or Just Your Keywords?

●       Part 18: Single-Channel AI in a Multi-Channel World  — This article you just read

About This Series & The Work Behind It

This AI Realities series is the published layer of a larger mission — helping professionals, trainers, and organisations navigate structured AI adoption with clarity and confidence. One pattern has emerged consistently over four years of hands-on AI work: most teams focus on getting better outputs, but very few understand what the model fundamentally cannot do — and what that means for how they must show up alongside it.

📘 AI for the Rest of Us and related practitioner guides — available on Amazon — move from foundational principles to structured application frameworks for professionals and business leaders.

💼 As a management consultant and AI strategy partner, work with organisations spans AI governance, workflow design, leadership training, and structured AI adoption programmes — not just demonstrations, but durable operating models.

Let's Stay Connected

Website & Blog: radhaconsultancy.blogspot.com

Contact Form: Contact through the blog form

Connect on social: LinkedIn  |  Twitter  |  Instagram  |  Facebook  |  YouTube – Radha Consultancy Channel

WhatsApp / Phone: Contact through the blog form (for consulting and training inquiries)

 Disclosure: This article reflects the author's interpretation of AI behavior based on professional practice and publicly available research. Political examples are generalised and unnamed by design. Created with AI assistance under strict human supervision. Information accurate as of August 2026. Verify independently for critical decisions.

#AIRealities #AILiteracy #InformationArchitecture #FutureOfWork #CriticalThinking #PromptEngineering #ManagementConsulting #ResponsibleAI