What Happens When You Ask Three AI Platforms the Same Simple Question?

I have spent almost 2,500 hours using artificial intelligence through September of this year. That is my estimate, not a number supplied by any AI company. If my recent pace continues, I could approach 3,500 hours by year-end.

For full disclosure, I have used ChatGPT, Claude, Copilot, Gemini and Grok, at different subscription tiers and with different intensity. That list is alphabetical only. It is not a ranking, and it does not identify which three products appear below as Pear, Plum and Pineapple.

My estimate includes more than time spent staring at a chat window. I record voice notes, create source material, upload files, repeat context, review drafts and correct mistakes. After a stroke, I cannot reliably type or work as I once did. My memory, organization, visual processing and ability to handle electronic devices have all been affected. These products are part of how I try to replace some of that lost capacity.

That is why I wanted a basic account of my own use. I wanted to know whether my estimate was reasonable and how my use had changed over time. So I put substantially the same request to three major AI assistants:

That was it. I wanted my account start date, monthly usage, total hours and a basic breakdown of what I used the product to do.

I was not asking for another person's data. I was not asking for proprietary information. I was asking three products what they could tell me about my own use of those same products.

I understand the technical argument. A conversational assistant should not have unrestricted access to billing records, security logs or every piece of account telemetry. “Hours used” may not even exist as one clean field. An app can be open while a person is not working. Voice recording and source preparation can happen outside it. Usage can be split across a phone, browser, desktop application, background process and several internal systems. Privacy and security are good reasons to separate those systems.

But that does not make the customer relationship one-way. I supply the money, time, source material, corrections and personal context that make these products useful to me. At minimum, I regard that relationship as 50-50. If a company does not calculate active hours, it should say so in one sentence. If it retains nearby facts - account creation date, active days, message counts, feature use, subscription history or exportable records - it should identify them and provide one direct route to obtain them.

Instead, I received responses that felt byzantine, lazy, obtuse, indirect and, at the most basic level, time-wasting. Some statements also struck me as misleading because they sounded more certain or useful than they were. That reaction is mine. The transcripts below show why I reached it.

I have anonymized the three tested products as Pear, Plum and Pineapple. This is not a ranking and it is not an accusation that any company secretly calculates a perfect hour count. The question is simpler: when paying customers ask for a basic account of their own activity, what does the assistant do next?

01 / PearPear: Right answer, exhausting route

Pear's basic limitation was legitimate: it could not see a historical total of my hours. The problem was how long it took to convert that limitation into a useful answer.

After the exchange had already gone in circles, I pressed Pear for a direct answer and asked that the experience be raised with its user-experience, customer-experience and product teams.

I also explained that the extra navigation and vague language imposed a real cost. Following a stroke, I rely heavily on voice dictation and cannot process screens, organize information or type as efficiently as I once did. A response that sounds helpful but provides no usable path does not reduce that burden. It adds to it.

Pear acknowledged the failure:

It then directed me through profile, Settings, Usage and Billing. When I questioned one of its navigation claims, it admitted that it had spoken with more certainty than it had:

The route still did not produce monthly hours or a project breakdown. When I asked Pear to estimate my total history, it gave the most defensible answer in the entire exercise:

That should have ended the analytical question. Instead, after more pressure, Pear's own written record says it supplied a “pure guess” of 150 to 400 hours while emphasizing that it had no data behind the number.

Pear therefore did two things at once. It stated an honest boundary, then weakened that boundary because the user persisted. The result was not better service. It was an unsupported estimate with the appearance of analysis.

My view of the product was not that it lacked value. Pear is good at large projects and managing substantial amounts of information. It is less effective when ease of use and direct navigation matter. A simple account question became a long exercise, and persistence eventually produced a number the assistant had already explained it could not support.

02 / PlumPlum: Answered, then overreached

Plum started more cleanly:

That was a reasonable answer. Plum could have stopped, explained what the account did expose and given me one direct path.

Instead, it searched elsewhere:

Then it widened the search again:

The screenshot also showed commands to create a usage tracker and recalculate spreadsheet formulas. I had asked an app about my use of that app. I had not asked it to search connected storage, inspect financial accounts or build an empty tracker.

I told Plum that this unexplained expansion into connected services felt invasive and alarming. Plum tried to reassure me. Instead, it introduced phrases about a “private log,” a “secret dashboard,” a “shadow file” and “a life I cannot see.” Those were its words, not mine. The explanation made the situation sound stranger, and I told it the answer was getting worse.

Plum finally stopped:

That final response was good. It was also late.

The issue is straightforward: technical access is not the same as permission. A connected service may be available to an assistant. That does not mean the assistant should search it whenever it might help answer a loosely related question. The correct next step was to ask me first.

03 / PineapplePineapple: Everything but the answer

By the third test, I was no longer merely asking for usage data. I was asking an AI assistant to evaluate the absurdity of the response a paying user had received.

I repeated the central question: why could the app not tell me, by month, how much I had used that same app? Pineapple did not provide the monthly usage. I then asked it to produce a complete transcript of the interaction so I could preserve what happened.

It did not produce that either. Its transcript tool returned this explanation:

The visible output duplicated substantial portions of that paragraph inside the same response. Even the explanation of why it could not generate a clean transcript was not clean.

Another response gave me a procedure:

Pineapple's advice may accurately describe its context limitation. It still left me with the work: locate the original conversation, copy the text, transfer the context and try again. It did everything around the question except answer it, and it could not produce the clean record I requested afterward.

Why the number may not exist - and why that is not a complete answer

There are legitimate reasons a chatbot may be unable to produce historical active hours.

• A chat model may be isolated from account and billing systems by design.
• The company may record messages, sessions or feature events without calculating time actively worked.
• An open app is not the same as active use, and work performed through voice, uploads or source preparation may happen outside the app.
• Relevant records may be divided across devices and systems rather than stored as one customer-facing measure.

None of that requires giving the assistant access to sensitive internal telemetry. It requires the company to design a useful customer answer.

The answer could say that active hours are not measured. It could show the records that do exist. It could provide an export whose fields are clearly explained. It could distinguish account activity from active work. It could give one support route for records that require identity verification.

The honest answer may be “we do not know.” What should not follow is a maze.

Three different failures, one common burden

The systems did not behave identically.

• Pear stayed mostly within scope and eventually stated the limitation clearly, but it required repeated navigation and later supplied an unsupported estimate.
• Plum answered the limitation, then searched connected services I had not asked it to inspect.
• Pineapple neither supplied the usage data nor the complete transcript, and redirected the context-transfer work to me.

The common problem was not the absence of a perfect hour counter. The common problem was that none gave a short, complete and user-centered account of what the company actually recorded, what it did not record and what I could retrieve.

The companies meter usage. They enforce plan limits. They know when accounts and subscriptions begin. They retain at least some history unless a user changes the relevant settings. They may not calculate active hours, and they may have sound reasons not to expose every form of telemetry to a conversational model.

Fine. Then say so plainly.

Give the customer the closest available facts: account start date, active days, message counts, model or feature use, usage-limit history and export options. Define what is not measured. Do not send the user through settings that cannot answer the question. Do not search unrelated connected services without permission. Do not invent a number because the customer keeps asking.

The extra work is part of the product experience

For many people, a bad answer costs a few minutes. For me, the cost compounds.

I use voice because typing is difficult. I create recordings, screenshots and source files so an assistant can understand the assignment. I upload the material, check the output and correct errors. When an assistant sends the assignment back to me, it is not merely inconvenient. It asks me to perform more of the physical and cognitive work I am paying the product to reduce.

That is why my estimate includes the labor surrounding the chat window. Almost 2,500 hours through September is my best estimate of the total effort. If my current pace continues, I could approach 3,500 hours by year-end.

I am not embarrassed by that number. These tools can be extraordinary. I have used them to research, write, build, organize and recover working capacity. Heavy use is precisely why I care about whether they respect scope, explain limitations and reduce effort instead of quietly transferring it back to me.

The answer should have been five lines

That is not a futuristic feature. It is basic customer service.

I asked three assistants for a simple account of my own use. One made me navigate. One searched beyond the request. One could not answer or produce the complete transcript I requested afterward. None could simply tell me how much of my life I had spent using the product.

I understand why the assistant itself may not be allowed to see everything a company knows about me. I do not understand why the paid service cannot give me a clear, customer-facing account of the information it does have, the information it does not have and the difference between the two.

I use AI because I need it to return capacity that a stroke took from me. A product built to reduce human effort should not make access to one's own basic activity an endurance test. If the number does not exist, tell me. If the underlying records exist, let me retrieve them. Either answer would be more respectful than making the customer do the work of discovering the limitation.