The more I use AI chatbots, the more questions I have.

AI assistants are quickly becoming part of many everyday workflows. We use them to write, summarize, translate, brainstorm, debug code, explain concepts, and sometimes simply to think through a problem, but usefulness and responsibility are two different things.

Using a tool responsibly means understanding what you are using and how you are using it.

Behind an interaction with an AI chatbot there is infrastructure, which I wouldn’t really define as “lightweight”. There are servers, data centers, storage systems, privacy policies, legal obligations, energy consumption, and training pipelines.

None of this means AI should not be used, which could be almost impossible in some case given the technological advancement of the world, but I do think responsible use starts with asking more questions about the systems we interact with.

For example:

• what resources does it require?
• what happens to the data I share?
• when should I use it, and when should I not?

Let’s unpack some of these questions here.

The environmental question

There is a lot of conversation happening on LinkedIn, in forums, and across communities of different kinds about the environmental impact of large language models.

This is not a simple topic, and may I add, numbers and fear-mongering claims are often shared without context. The underlying point, however, is very very real: AI systems require significant computational infrastructure.

What does computation require? Energy. A lot of it.

AI chatbots run on large data centers filled with specialized hardware designed to process massive amounts of data. These facilities consume electricity for computation and for cooling systems that prevent the hardware from overheating. In some cases, water is also used to help cool servers.

I know what you’re thinking now, how much energy does my prompt consume?

Energy per query is variable, model and workload dependent, and strongly influenced by output length, system utilisation, and data centre efficiency. Long, multi-step, agentic workflows are not the same as “what is the capital of Denmark?”.

To put it into perspective, a query on an AI chatbot consumes in general more energy than a normal search on Google, which makes sense because the two systems operate in different ways.

A search engine retrieves indexed pages. A genAI chatbot, on the other hand, generates text token by token using a neural network that performs billions of calculations during inference. This additional computation increases the energy required for each request.

At a global level, the impact becomes more significant because of scale. Data centers already consume roughly 1.5% of global electricity.

The bad news is that demand is expected to increase as AI adoption grows. The good one is that with progress and through research innovations, software and hardware efficiency improvements, comes optimization of how much AI systems consume.

However, this topic is often framed in overly simplistic ways (here as well of course). A single prompt represents a very small unit of energy consumption and for most individuals, it accounts for only a tiny fraction of their overall digital footprint. The real impact comes from aggregate usage across billions of interactions, not from a single person asking a question.

So the point here is not to feel guilty about using AI. Chances are it is not going anywhere, on the contrary, usage will likely continue to grow.

The point is awareness.

Just as we eventually learned that streaming video, cryptocurrency mining, and cloud storage have environmental costs, AI too belongs to the category of energy-intensive digital infrastructure.

To me, responsible use means recognizing that computation is not free. That awareness can, and should, influence how we choose which tools we use and how.

The data question

Data today is one of the most valuable resources in the digital economy. It fuels pretty much everything, advertising systems, engines, platforms, and yes, AI models too.

When you use an AI chatbot, you are not only asking and receiving output, you are also providing input. That input can be valuable, sensitive, regulated, or simply more revealing than you intended.

Every prompt you send is a piece of data that enters a system operated by a company with its own infrastructure, policies, and legal obligations. This does not automatically mean something unsafe is happening, all the AI providers I know have clear policies about how conversations are handled, however, the details vary widely between platforms and plans.

The issue here is that most users never read them.

Different layers of data handling may exist behind the interface, including:

• conversation storage
• security monitoring
• model improvement pipelines
• legal compliance requirements
• vendor processing infrastructure

Because of this, there are a few basic questions I think are worth asking when using AI systems:

• are my conversations stored?
• are they used for training?
• can human reviewers access them?
• what happens when I delete them?
• does the policy change depending on the plan I use?

Consumer side

For consumer services, OpenAI’s privacy policy states it may use content submitted to its individual services (including prompts, responses, and also images/files) to improve model performance, depending on user settings.

It also states that a limited number of authorised personnel and trusted service providers may access user content under specific conditions (support, abuse/security investigations, legal matters, or model improvement unless you have opted out), and it explicitly advises users not to enter sensitive information they would not want reviewed or used.

On deletion, OpenAI explains that cleared chats are deleted from systems within 30 days unless they were already de-identified or must be retained for security or legal reasons.

If you want a more “one-off” mode, OpenAI’s Temporary Chats are deleted after 30 days, are not used to train models, and may be reviewed only for abuse monitoring.

Anthropic’s privacy policy explains something similar but not identical: you can delete conversations (removed immediately from history and deleted from back-end storage within 30 days), but if you allow your chats to be used to improve the model, data may be retained in de-identified form for up to five years in training pipelines. It also lists longer retention for certain flagged policy-violation cases.

Meanwhile, Google’s Gemini Apps Privacy Hub is very explicit about retention edge cases: it notes that you can change auto-delete settings and manually delete chats, but chats reviewed by human reviewers are not deleted when you delete your activity, and can be retained for up to three years.

It also states that even when the “Keep Activity” setting is off, Google still uses chats to respond and to help protect users and the public (including with help from human reviewers), and that “temporary chats” or chats with “Keep Activity” off are retained with your account for 72 hours.

These policies are designed to make the chatbots safer, but they also mean that you should not treat a chatbot like a diary or your best friend.

Enterprise side

Some business plans change the default settings.

OpenAI states that for business services (ChatGPT Business/Enterprise and its API platform), it does not train on inputs or outputs by default, and that organisations are opted out of data sharing unless they explicitly opt in.

Its Services Agreement likewise says it will not use customer content to develop or improve the services unless the customer explicitly agrees.

In the Google ecosystem, the “Generative AI in Google Workspace” Privacy Hub states that prompts and content are not human reviewed or used for model training outside the customer’s domain without permission, and that admins can control retention windows.

In the Microsoft ecosystem, Microsoft states for Microsoft 365 Copilot that prompts, responses, and data accessed through Microsoft Graph are not used to train foundation models, and that the system operates within the Microsoft 365 service boundary with existing compliance commitments for commercial customers.

The info is there!

Still, none of this is obvious unless you actively look for it.

Responsible use starts with a few simple things:

• knowing which plan you are using,
• reviewing privacy settings,
• understanding data retention policies,
• avoiding the assumption that deletion automatically means disappearance.

At work, this awareness becomes even more important.

If a document is considered confidential inside your company, pasting it directly into a consumer AI tool should not automatically be the default choice. Understanding where data goes is part of using any digital infrastructure responsibly. AI systems are no exception.

What about the uploads?

Uploading a picture is so easy, right? To me it doesn’t feel that easy when I know that pictures can contain faces, locations, and context that are automatically shared with the system when you upload them. The same applies to any files like PDFs, spreadsheets, and presentations.

When you upload a file or image to a chatbot, you are typically uploading it into the same data-handling system as your conversation.

OpenAI explicitly includes “images and files” in the category of content that may be used to improve model performance for consumer services, depending on settings.

It also states that files uploaded during a conversation are tied to the conversation lifecycle: delete the conversation, and the file is scheduled for deletion alongside it (with the same 30-day window and the same legal/security exceptions).

Google’s Gemini Apps Privacy Hub is similarly explicit that “information you provide” can include prompts and uploaded content, and it highlights that human review and retention rules can apply to a subset of chats and related data.

If the task does not truly require the original file, do not upload the original file. If it does, reduce the information first.

Good rules of thumb

• Crop screenshots so they show only what matters,
• remove or avoid location metadata when sharing personal photos,
• redact names, IDs, and internal URLs before uploading,
• and, at work, use enterprise configurations for proprietary material instead of consumer tools.

Necessary vs convenience compute

This is the line I keep coming back to with my questions. There is necessary compute, and then there is convenience compute.

Necessary compute is when the tool unlocks something you could not realistically do otherwise: compressing a week of research into a structured brief, brainstorming angles you can refine, generating a draft you will heavily edit, translating or analysing at scale, supporting accessibility.

Convenience compute is when the tool replaces minor friction: rewriting a sentence ten times because, asking the model things you could have retrieved with a direct source, turning every tiny decision into a prompt because it feels easier than thinking for thirty seconds.

The line that gets drawn is personal and context dependent, but surely asking the question already makes you use the tool differently.

It nudges you to:

• reduce needless iterations (one well-scoped prompt beats ten vague ones),
• choose smaller or more efficient options when the task is simple,
• and reserve heavy “reasoning” flows for when you actually need them.

What does responsible use actually mean?

My point of view is that responsible use is a set of defaults you can live with.

• I do not default to AI for everything.
• I check privacy settings once, properly, and I revisit them when tools change. If you want a practical starting point for that habit, read my guide on how to navigate ChatGPT settings.
• I avoid uploading sensitive material casually, because uploads feel temporary, but the systems behind them are designed for retention, review, and compliance.
• I treat important outputs as drafts, not truths, and I verify what matters.
• When the task is serious, I treat the tool like infrastructure, not magic.
• I try to stay curious and keep asking the questions, keep learning and stay updated.

Let me say this again: AI is not magic. AI does not “think” as they write on the loading screen. AI is a tool and it needs to be used as such if we want to make the most out of it.

Other resources I found very interesting about this topic:

• Search with Candour episode on How we’re using AI in our work processes (feat. Sarah McDowell)
• How to effectively govern AI use in content teams by Sarah Mitchell, SALT Agency
• We did the math on AI’s energy footprint. Here’s the story you haven’t heard. MIT technology Review