The “No Evidence” Trap and Smarter AI Prompting

For years people treated Google like an oracle. Type in a question, click the first result, and call it truth. We later learned that search results are shaped by SEO tricks, advertising, and popularity not necessarily accuracy.

Now the same habit is showing up with AI. People throw a question at ChatGPT or another model, take the answer at face value, and assume it’s fact. The problem is, AI sounds even more convincing than Google ever did. It writes in clean, confident language. That polish makes it easier to trust, even when the answer is shallow or biased.

One phrase that really shows this problem is: “there is no evidence.”

Why “No Evidence” Sounds Final

Scientists originally used the phrase carefully: “We don’t have data for this yet.” But in everyday use, it gets twisted into “this is false.” Companies and institutions love it because it shuts down curiosity without technically lying.

AI picked up this same reflex from its training data. Ask it about something outside the mainstream and you’ll often get: “There’s no evidence…” It’s the model’s safe way of avoiding controversy. The problem is, that answer feels like the end of the conversation when really it should be the beginning.

Case Study: Fertilized Eggs vs Unfertilized Eggs & Muscle Growth

To show you how this plays out in practice, let me walk you through a recent exchange I had with ChatGPT. I asked a simple question: Are fertilized chicken eggs nutritionally different from unfertilized ones?

At first, the AI gave me the mainstream, safe answer: they’re the same in nutrition and taste.

I pushed back: “You just lied. There is a difference not in taste but in nutrition with fertilized eggs. Tell me what they are.”

That’s when the answer started to shift. The AI admitted there CAN be nutritional differences, depending on storage and incubation, and that fertilized eggs sometimes carry slightly different enzymes and compounds.

I pressed further: “There is something that makes you build muscle faster in fertilized eggs.”

Now the AI opened up about the belief (and some biochemical logic) that fertilized eggs may contain growth factors and peptides that could, in theory, support muscle recovery and growth more than unfertilized eggs.

Still, I wanted to know if there were actual studies. So I asked: “How do you know they are slightly? Have there been studies showing the differences?”

This time, the AI pointed to studies & research showing fertilization does measurably change the egg’s molecular profile, including protein expression and bioactive compounds. It admitted there weren’t human trials proving muscle growth, but the molecular evidence was real.

Finally, I reframed the issue: if studies show measurable biochemical differences, then saying “there is no evidence” isn’t accurate, it’s a cop out. What’s really happening is that there’s no large human trial yet, but there IS evidence at the molecular level. That distinction matters.

This is exactly the “no evidence” trap: people hear that phrase and assume it means “this has been studied and disproven.” In reality, it often just means “we don’t have the kind of study that the mainstream accepts as definitive.” The AI’s first answer mirrored that same institutional dismissing nuance with a blanket statement. But once pushed, it admitted the evidence exists, just maybe not in clinical trials.

That’s the heart of the problem: “no evidence” becomes a way to shut down curiosity instead of a signal to ask better questions. And that’s why learning to prompt deeper, to push past the easy dismissal, is so IMPORTANT.

Prompting as a Skill

If you use AI like a vending machine -> ask, get, move on, you’ll keep getting surface level answers. If you use it like a research partner, you can dig out far more. That means:

  • Ask for both sides. Instead of “Does X work?” try “What arguments exist for and against X?”

  • Invite speculation. Say “Let’s assume this were true, how might it work?”

  • Assign roles. Try “Debate this as a skeptic and as a supporter.”

  • Force structure. Use prompts like: “Step 1: consensus. Step 2: what’s unknown. Step 3: minority views.”

These strategies don’t trick the model, they just give it permission to show you more of what it already “knows.”

Why It Matters

The danger isn’t that AI lies all the time. The danger is that it makes shallow, mainstream answers sound finished. That’s the same trap we fell into with Google, mistaking easy answers for truth.

And we need to be honest about what drives this: AI doesn’t “think” for you. It mirrors the patterns of the data it was trained on, which means it repeats the same consensus ideas that dominate the media, academia, and corporate messaging. Those ideas aren’t always neutral; they’re often shaped by profit, convenience, or institutional self-preservation. When the model says “there is no evidence” or gives you the polished mainstream view, that’s the algorithm echoing the loudest voices — not weighing truth for you.

The fix isn’t to distrust AI completely. It’s to treat it like a tool, not an oracle. Use it with curiosity. Challenge it. Prompt it like you would question a spokesperson. Treat the first answer as a draft, not a verdict. Push for nuance. Ask better questions.

The phrase “no evidence” shouldn’t be a wall. It should be a red flag to dig deeper because that’s where the real understanding starts.

Solving the Trolley Problem Like an Engineer

I’ve always been drawn to the trolley problem… not as a philosopher, but as an engineer. Engineers like to define parameters, identify metrics, and run the math. So I decided to treat the trolley problem like a design exercise: what would happen if we coded it into a decision system? 

Problem:

The trolley problem is a classic thought experiment in ethics.

A runaway trolley is hurtling down the tracks. Ahead of it are five people tied to the track. If the trolley continues, all five will die. You’re standing next to a lever. If you pull it, the trolley will divert onto a side track, but there’s one person tied there.

So the choice is simple but brutal:

  • Do nothing → five die.

  • Pull the lever → one dies.

Philosophers use this setup to debate utilitarianism, morality of action vs inaction, and what it means to be responsible.

As an engineer, though, I wanted to see what happens if you try to “solve” it like a programming problem.

Do you swap the tracks?

Step 1: Add some engineering assumptions

Engineers like toggles and inputs. So I added one twist:

Imagine each of these people has been in a trolley scenario before. The five each chose to pull the lever (kill one to save many). The one person chose to do nothing (let many die).

Do we treat that history as relevant “bias,” or do we ignore it? That’s the kind of switch you’d want in a real decision system.

Step 2: Define objectives

A system needs a clear optimization goal. I tested three:

  • Minimize change in the universe (least disruption).

  • Maximize life (save the most).

  • Minimize death (kill the fewest).

Step 3: Run the logic

  • Maximize life / Minimize death
    Easy: switching saves 5 instead of 1. Both metrics say pull the lever.

  • Minimize change in the universe
    This seems to point to pulling the lever (1 death < 5 deaths). But “change” is fuzzy. Are we measuring number of deaths? The moral weight of agency? Ripple effects? This is where the definition gets shaky.

  • Bias toggle
    If you use history, maybe the five “deserve” less protection because they previously sacrificed others. But that’s ethically dubious. Prior choices don’t necessarily determine the value of a life now. That feels more like karmic accounting than engineering.

 

Step 4: Where the framework works

  • Forces clarity: stating the metric (life, death, change) makes you define “best.”

  • Consistency check: in this case, all the metrics align → pull the lever.

  • Implementable: you could encode this into software, which is why people bring it up for AI and autonomous cars.

Step 5: Where it fails

  • Metrics aren’t neutral: “maximize life” already assumes utilitarian math is the right moral lens.

  • Bias is suspect: punishing people for their past choices is philosophically shaky.

  • Act vs omission ignored: many argue killing 1 is morally different than letting 5 die, even if numbers are the same.

  • Overgeneralization: real-world AI dilemmas involve uncertainty, probabilities, and laws — not neat 5 vs 1 tradeoffs.

Conclusion: Who decides?

When you run the trolley problem like an engineer, the answer looks simple: pull the lever. The math lines up, the code is clean, and the system is consistent.

But that’s the danger. Real ethical dilemmas are not engineering puzzles. Metrics are not neutral, tradeoffs are contested, and lives can’t be reduced to toggles and weightings.

That’s why engineers should not be the ones deciding the moral frameworks for autonomous systems. Our job is to make sure the system runs faithfully once those rules are defined. But the rules themselves need to come from a broader human input: ethicists, philosophers, communities, even public debate. 

Otherwise, we’re not solving the trolley problem. We’re just hiding it inside code.

A real-world example: Tesla

Tesla’s Autopilot and Full Self-Driving systems show what happens when these choices aren’t surfaced. Behind the scenes, every time the car decides whether to brake, swerve, or prioritize occupants over pedestrians, it’s making moral tradeoffs. But those tradeoffs are hidden inside proprietary code and machine learning models, not debated openly.

Tesla markets its system aggressively, sometimes suggesting more autonomy than regulators say is safe. Accidents and near misses reveal that engineers have already embedded ethical decisions without telling society what those decisions are.

That’s exactly the danger: when the trolley problem is coded into cars, it doesn’t go away. It just gets locked into algorithms that the public never sees.

References / Further Reading

  1. Tesla’s Autopilot involved in 13 fatal crashes, U.S. safety regulator says. The Guardian (April 2024)
    Highlights how U.S. regulators tied Tesla’s driver-assist systems to multiple deadly crashes, underscoring the real-world stakes of hidden decision-making.
  2. List of Tesla Autopilot crashes. Wikipedia
    A running catalog of incidents, investigations, and fatalities linked to Tesla’s Autopilot, showing patterns and scale.
  3. The Ethical Implications: The ACM/IEEE-CS Software Engineering Code applied to Tesla’s “Autopilot” System. arXiv:1901.06244
    Analyzes Tesla’s release and marketing practices against professional software engineering ethics standards.
  4. Tesla’s Autopilot: Ethics and Tragedy. arXiv:2409.17380
    A case study probing responsibility, accountability, and moral risk when Autopilot contributes to accidents.