Where AI is strong, and where it fails

In 2023, researchers from Harvard Business School, Wharton, MIT and others studied 758 consultants at Boston Consulting Group. On 18 tasks the AI could handle, people with AI finished 12.2% more tasks. They were 25.1% faster, and their work was over 40% better. One task was picked to be just outside what the AI could do. There, people with AI did worse. Out of every 100 people, about 19 fewer got it right. The researchers called this uneven edge a "jagged technological frontier." Read more in Centaurs and Cyborgs on the Jagged Frontier.

In a 2025 study by the research group METR, skilled software developers used early-2025 AI tools. They took 19% longer on their tasks. Yet they believed AI had made them about 20% faster. In February 2026, METR said newer tools probably help developers more now. But it called its own new data weak evidence. So the lesson is not "AI slows you down." It is that a feeling is not a measurement.

• Facts, quotes and sources. It may invent them. Open every source it gives. Or give it the document and ask it to quote from it.
• Sums. It predicts digits as text instead of doing math. So long sums can go wrong. Ask it to run code, or use a calculator.
• Counting letters. It reads text in chunks, not letter by letter. So "How many r's are in strawberry?" can go wrong. Newer models usually get it right.
• Recent events. Use web search, and check the dates on what it finds.
• Rare topics. It saw little text about them, so its guesses are weaker. Check original sources, or ask experts.
• Long documents and chats. It can miss details in the middle. Ask about one part at a time.
• Agreeing with you. It learned partly from people's ratings of answers. So it leans toward what you seem to want to hear. This is called Sycophancy. Ask neutral questions, and ask for the case against. See when AI makes things up.

• ✍️ Drafting emails, letters and outlines.
• 🔄 Rewriting for tone, length or plain words.
• 📝 Summarizing text you give it. This is far safer than asking it to remember a text.
• 🌍 Translating everyday text between common languages. Take more care with rare languages, and with legal or medical text.
• 💡 Brainstorming names, ideas and questions to ask.
• 🎓 Explaining an idea at your level, as many times as you need.
• 💻 Code for small programs and spreadsheet formulas. You can run it to check it.

A tool is an extra program the model asks the app to run. Then the model reads the result. See tool calling. As of September 2026, ChatGPT, Claude, Gemini and Le Chat all offer some way to run code. Most apps show when a tool ran, like a list of sites or the code.

Priya asked again: "Use code to add these up and show me the calculation." This time she got a total she could follow. Still, tools do not fix everything. The model picks what to search for. It can choose a poor source or misread a good one. And no tool stops it from agreeing with you.