Editorial Team A 2025 Reuters Institute study found a jaw-dropping 87% of us ran into gen AI misinformation in the past year alone. That number proves our collective research skills aren’t ready for this technology, especially when AI chatbots deliver answers with a completely unearned confidence. So how do we fight back and learn to verify what these tools tell us?
Key Takeaways
- Misinformation from gen AI hit 87% of internet users last year, per a 2025 Reuters Institute study.
- Just 1 in 5 people actually check chatbot answers against other sources.
- Biased training data is a huge source of factual mistakes and stereotypes in LLMs.
- You can get up to 30% more accuracy by telling the AI to cite sources and explain its thinking.
- Old-school research skills like lateral reading are still the best way to use chatbots safely.
The Alarming Verification Gap: Only 20% of Users Cross-Reference
An early 2026 Pew Research Center study put a number on a problem I see constantly with my content teams: only 20% of people bother to check chatbot output against independent sources. People just accept what the AI says. This gets even worse when a deadline is tight because the promise of AI’s speed makes it easy to skip verification, which is exactly how bad information gets passed around so quickly.
The fault here lies as much with the AI’s design as it does with us. Chatbots are built to sound friendly and sure of themselves, giving you what feels like a final answer instead of a clue for your own research. That helpful, conversational tone can lull you into a state where you stop questioning things. You have to remember the AI’s confidence means nothing about its accuracy. It’s a machine built to predict the next word in a sentence, and that’s it.
Bias in Training Data: A Root Cause of Factual Errors
Many of the AI’s mistakes are just garbage in, garbage out. The problem starts with the biased data they’re trained on. A 2025 Stanford University analysis showed that large language models don’t just repeat the historical and societal biases found online. They actually amplify them, which results in factual mistakes and recycled stereotypes. You can see this in practice when you ask about a historical figure from an underrepresented group and get a vague or wrong answer, while a query about a more famous person yields a detailed biography.
The AI’s logic is working perfectly. The issue is the source material. Think of it like training a student using only the books from a single, biased library, their worldview would be completely shaped by that collection, no matter how smart they are. AI models are just sophisticated mirrors reflecting the internet’s content, warts and all. If the data they’re fed is skewed, their output will be just as skewed. These systems are simply reflections of their training data.
Prompt Engineering for Accuracy: A 30% Improvement
How you ask a question can make a huge difference. Late 2025 research from places like the Massachusetts Institute of Technology (MIT) found you can boost a chatbot’s factual accuracy by up to 30% just by using better prompts. Instead of a simple question, try ordering the bot to “Cite your sources” or “Explain your reasoning step-by-step and give me URLs to the original data.” This simple instruction forces the AI to ground its answer in something verifiable instead of just spinning out plausible-sounding text.
My team has seen this work firsthand. The moment we started demanding source attribution and a clear trail for where the information came from, the quality of the AI’s output shot up. The goal is to make the AI show its work. This moves the output from just being plausible to being defensible and traceable. Yes, it adds one extra step to the prompting process, but it saves us a ton of time on fact-checking down the line.
The Persistence of Hallucinations: A Lingering Challenge
Even with all the progress, AI models still “hallucinate”, they invent facts that sound completely real. A 2026 report from Google DeepMind admits that while hallucinations are less frequent in newer models, the problem isn’t going away. These fictions are so dangerous because they’re woven smoothly into text that is otherwise correct, making them almost impossible to catch unless you’re actively looking for them.
I don’t buy the argument that we can just train this problem away with bigger models. The improvements are real, but the core design of these systems is to predict the next most likely word, not to know what’s true. This means a certain amount of confident-sounding nonsense is baked into their very architecture. Trusting the AI to fix itself is a fool’s errand. We’re always going to need a human in the loop to apply critical judgment.
The Indispensable Role of Lateral Reading in the AI Age
With so much AI content flooding the internet, old-school info lit skills like lateral reading are our best defense. Research from Stanford University’s Civic Online Reasoning project (whose work extends into 2026) proves that people who use lateral reading are far better at spotting bad info. The technique is simple: when you see a claim, you open a new browser tab and investigate the source itself, instead of just judging the page you’re on.
My first move when an AI gives me a fact is to read laterally. I take a key phrase or name from its answer and pop it into a search engine myself, looking for results from a place I trust like AP News or Reuters. For instance, if the AI says the Fed is forecasting 2.5% GDP growth, I’m immediately searching “Federal Reserve economic forecast 2.5% GDP” to see the original report or a story from a reputable wire service. Doing this gets you around the AI’s potential for error and takes you straight to a real source. It’s the single best habit you can build to protect yourself from AI-generated junk.
In a world saturated with AI, building real information literacy skills isn’t optional. You have to verify sources. Question everything the AI tells you, cross-reference its claims, and trust your own independent verification, not its confident tone.
What is chatbot accuracy?
It’s the measure of how correct, verifiable, and error-free the information generated by an AI chatbot is. High accuracy is a direct reflection of the AI’s reliability.
Why do chatbots make factual errors?
They make mistakes for a few reasons: the training data is full of human biases, their design is based on predicting the next word (not on checking facts), and they sometimes “hallucinate,” which means they just make things up.
How can I improve a chatbot’s accuracy when asking questions?
Use more specific prompts. Tell the AI to “cite its sources,” “explain its reasoning step-by-step,” or “provide URLs to authoritative data” to force a more accurate, grounded response.
What is lateral reading and why is it important for verifying AI information?
It’s a research technique where you open new browser tabs to investigate a claim or source, rather than just trusting the page you’re on. This is the best way to verify AI output because it lets you cross-reference claims against independent, trusted sources and bypass the AI’s own biases or mistakes.
Can AI tools ever be fully trusted for factual information without human verification?
No. They should always be considered a starting point, not a final authority. Because they can still produce errors, reflect biases, or fabricate information, human critical thinking and verification are still absolutely necessary to confirm facts.
