# Research Tools --- ## AI assistants — AIAs I use a variety of common tools for conducting my research, organizing my findings, and writing. Most of them aren’t worth mentioning — with one exception: AIAs, such as ChatGPT, Claude, and Gemini. Note: Tools are not methods. For the latter (how I research) see [[Methods|Research Methods]]. ## When and how I use AI in my research Since AI is the primary focus of my research and client work, I’m acutely aware its strengths and shortcomings — and therefore of the trade-offs involved. That’s why: - **I don’t use AI to write on my behalf.** - **I use AI for various other tasks — taking care to sidestep the risks.** Here’s why, with some high-level explanation of risks and remedies. ### Not for writing AIAs write adequately for some purposes — better than most humans, in fact.[^1] But I appreciate it when materials I’m *reading* are better than what an AIA can generate. So I want to return the favor by holding myself to that higher standard when I’m *writing*. That’s why I write all my research briefs myself; the same is true when I write other material, for clients. ### How, for other tasks For tasks other than writing, I use AIAs daily. But I know their weaknesses well — so I always take measures to sidestep them, specific to each category of work.[^2] Here’s a high-level overview of the main tasks I use AI for, why I use them, the potential pitfalls, and how I sidestep them through prevention, abstention, and correction: | Task | Benefit | Risks and remedies | | ----------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Summarize** | Reading AIA-generated summaries of lengthy PDFs and other types of sources helps me decide which ones are relevant to me, in whole or in part — and, if in part, which parts. | The summaries an AIA generates occasionally distort the source, so I never fully rely on them. If, based on a summary, it seems the source is relevant, I always read the original myself. | | **Search** | AIAs do better than traditional search at locating results based on the *meaning* of a prompt rather than the verbatim words in a search query. | Sometimes AIAs *generate* links instead of listing links to sources they’ve retrieved. So I include an AIA-provided link in a citation only after reading the source it points to myself, to confirm its relevance. | | **Study** | AIAs often explain topics in depth better than any individual source can, and provide links to multiple sources to bolster the credibility of their explanations. | AIAs sometimes conflate disparate sources, blending them in ways that result in inaccuracies. So before trusting an explanation, I follow the links to sources to verify it. | | **Brainstorm** | Brainstorming with the right people often yields deeper insights than with AIAs — but I love brainstorming with AIAs too because they’re better than most people at spanning disciplines. | AIAs sometimes veer into territory that isn’t relevant to my purposes, so I course-correct to steer them in the directions I want to explore. | | **Pressure-test** | AIAs can spot flaws in argumentation when I ask them to identify erroneous logic, false premises, or leaps in reasoning in my drafts. | I overcome the problem that AIAs are designed to be pleasant (which weakens their critiques) by explicitly instructing them to push back and to argue “the other side.” | | **Polish** | LLMs contain rich representations of subtle relationships between meanings of words and phrases. So even as an experienced author, I find AIAs to be useful partners in refining my language. | Although AIAs often demonstrate uncanny mastery of nuances of meaning and tone, they’re sometimes just wrong. I trust my judgment as the tie-breaker. | ## The exceptions In light of the benefits of using AIAs, it would be inefficient not to use them for certain tasks (with appropriate precautions). But there are moments when efficiency takes a back seat and, if time allows, I choose to do what an AIA could do quite well myself — because work is also about the pleasure of the process, not just the satisfaction of the outcome. [^1]: Occasionally, in my interactions with AIAs, they’ve generated turns of phrase that are beautifully worded and subtle in tone. It’s rare but, when it happens, I love it and find it fascinating. [^2]: It’s instructive to compare how AIAs perform to how humans do, to reflect on their relative strengths and weaknesses. During my years at Forrester Research, as an analyst and in research leadership, I collaborated with colleagues in each of these categories of work, of course. Today’s AIAs can at times be better than I and my Forrester colleagues (or any human, for that matter) are at certain things, weaker at others. However, there are two practical areas where AIAs stand out. They’re available anytime and anywhere, which is convenient. And they never get tired or distracted, which makes a difference in the quality of work because of the value of iteration.