Protecting ourselves from 2 key risks of AI tool dependency

In Part 1 of this 2-article series, I compared reliance on LLMs for our daily lives or livelihoods to wearing eyeglasses. Here in Part 2, I cover what we can do about two risks of being dependent on AI tools: cognitive effects and tool access disruption. If you’re not dependent on AI, I invite you to share it with someone who is! (If your email client truncates this, click here to read the full article.)

Here’s Part 1, in case you haven’t seen it yet:

Technology

Eyeglasses and LLMs considered harmful?

·

Jun 17

Today I’m thinking out loud a bit about analogies in which AI tools are compared to everyday assistive devices like glasses, hearing aids, crutches, knee braces, wheelchairs – or to sci-fi devices like Iron Man exoskeletons.

Regardless of the reasons why someone depends on an LLM, key risks loom. Today we’ll look at two risks from dependency on an AI tool: cognitive effects and tool access disruption.

There are other risks, such as not fully owning what you generate with an LLM. See Everyday Ethical AI for more on AI risks, or Sena Evren’s article for more insights on IP rights in particular.1

Unlike wearing eyeglasses, which doesn’t diminish natural eyesight, over-reliance on an AI tool can create risks of losing some of our cognitive abilities.234

That makes LLMs more similar to assistive devices such as knee braces. Braces can help people with bad knees stay mobile, but over time they can also cause weakness in the leg muscles. (More info on bad knees and braces here5.)

I talk a lot with my AI, Software, & Wetware interview guests about keeping our biological human brains (‘wetware’) engaged when using AI. This isn’t new with AI, but it is too easy to allow a technology to shift from being an assistive tool to becoming a dependency. (Example: calculators.) It’s even easier with AI because many tools are designed to keep us chatting, spending money, and becoming dependent. (Calculators don’t do that.)

That doesn’t mean we shouldn’t make the shift and use AI, just that it should be a mindful, deliberate, controlled choice. We need to think about what cognitive ‘muscles’ we might be allowing to atrophy by making specific choices on how and when to use AI. And thinking about these tradeoffs may prompt us to look for other ways to keep those cognitive muscles strong, while still using AI tools to give us a boost where it adds value for us.

People have gone back and forth about various studies showing that using AI rots our brains, or doesn’t. I wrote about this a bit in my Everyday Ethical AI book. The biggest factor seems to be not whether we use it, but how we use it. Some examples:

• Asking an LLM a question, vs. asking it to ask us challenging questions.

• Accepting an answer from a LLM as accurate or factual, vs. challenging its answer to surface and correct biases or errors.

• Asking the LLM to write for us, vs. asking it to critique our writing and point out holes in our arguments or bugs in our code.

• Asking the LLM to summarize a paper for us, vs. reading the paper and then asking it questions about interpreting what we read.

• Using an LLM to provide companionship in lieu of humans, vs asking it to help us communicate more effectively with the humans in our lives.

Think of how we use LLMs as the difference between training a muscle with good form vs. bad form. Training with good form builds strength. Training with bad form can cause injuries (to ourselves or others). Of course, training with "good form" can look different depending on the person and their life circumstances or even an individual’s day-to-day capabilities.

Exercising our brains in how we use AI is like learning ways to train a bad leg at home, while still wearing a knee brace to stay safe & mobile when going out, to prevent loss of strength in the leg. We face a similar risk to our brains with AI. “If you don’t use it, you lose it” is a common saying because it’s true.

For people who now rely on AI tools for how they live or how they make a living, regardless of the reason, AI tool dependency is an important tradeoff.

The risks of suddenly being without an AI tool that someone’s become dependent on are much higher than me breaking or losing a pair of glasses, though.

• Any software company could be hacked or disrupted at any time, and that includes AI tool providers (e.g., the AWS, Claude, and Vercel outage on March 2, 20266).

• Even if your internet connection is normally reliable, and no one hacks or breaks the AI tool, the tool still may not be reliable. A model update you don’t control could change model behavior in a way that breaks a feature you rely upon.7

• Tool providers may announce massive price increases, causing the costs could suddenly change in a way that breaks your business model. (Not a risk with glasses; once I buy them, they’re mine to keep.)

  • See Harsha Maddipatla’s “Token Price Index” tracker for a view of recent changes; for instance, on May 29 he reported that Google tripled Fast tier pricing.8

• Political or commercial influences may cause a tool you rely on to suddenly be taken offline. For instance:

  • The Jan. 29, 2026 announcement by OpenAI that they were again retiring GPT-4* models (after the August 2025 outcry the first time they did it).

  • The June 12, 2026 announcement from Anthropic that they were taking Mythos 5 and Fable 5 offline worldwide due to US government pressure.

For people who live in parts of the world where power or internet connectivity are unreliable, the day-to-day risks of losing AI tool access are even higher.

Disruption of AI tool availability is a business and personal risk. How big of a risk? That depends on how heavily you rely on the AI tool. If your AI tool use is more like donning an Iron Man powered exoskeleton than wearing a knee brace or eyeglasses, then a tool disruption is going to have dramatically bigger effects.

First, we can keep our wetware engaged and active.

In recent years, dentists in my area have adopted the slogan “You only need to brush and floss the teeth you want to keep.” A bit scare-monger-ish, but effective at getting the point across. Likewise, we only need to exercise the cognitive skills we want to keep, such as:

• Do math in your head or on paper once in a while, enough to help you check if a calculator or spreadsheet or AI-generated financial projection is right.

• Develop a habit and method for checking if an AI tool’s response contains errors or reflects hidden biases, and correcting them.

• Keep learning.

Note that learning or training our brains is likely to look different for everyone, and may vary for an individual from one day to the next.

For ideas on keeping our wetware fully engaged in how we use AI tools, check out writers here such as Dr Sam Illingworth (Slow AI; see his article on AI deskilling), Mariam Vossough (The Women’s AI Voice; see her article on why your AI drafts don’t feel like yours); and many others. The Business section of the She Writes AI directory I curate is one of our largest categories. My co-authors in the SheWritesAI collaborative books also write about using AI tools judiciously in our personal lives, including:

• AllAboutAI by Midhat Tilawat on using AI as a helper, not a hand to hold forever for decision-making,

• Andrea Hiott on how LLMs lead us to care (more love, less manipulation),

• Sarah Ennett - Seeking Ikigai on the ikigai-risks of AI.

More chapters on keeping & building our humanity and skills are coming this fall in the second SheWritesAI book collaboration, such as Jax on reclaiming our agency with AI, and Devika Toprani on how Somagraphic Learning™ puts human cognition back in the driver's seat. You can also find their writing now in the SheWritesAI weekly digests.

Second, for tool disruption risks, I encourage folks who depend on AI in their work or home life to think through a risk management plan covering both contingencies and mitigations. (In project management terms, mitigation = reducing the likelihood of something bad happening; contingency = reducing the impact when something bad does happen.) Disruption in AI tool availability is a risk, and one that could be personally stressful to navigate if it’s ever realized.

My risk management plan for my dependency on my glasses includes protecting the eyesight I have and putting my glasses in cases when I’m not wearing them (mitigations) and keeping spare pairs of old glasses handy (contingencies).

When I needed a knee brace, I kept a second brace and a pair of crutches handy as backup in case the first brace failed or my knee suddenly got worse.

For AI tool dependencies: have a backup plan, a backup tool, or a backup manual approach that lets you still do what’s most important if the AI tool(s) you depend on become unavailable. Think about two aspects of backup plans: loss of availability or access (which may include unaffordability), and unilateral or catastrophic changes in functionality. Your AI tool risk management plan might include tactics like:

• An alternative AI tool or gateway (account established and configured). Specific guidance on setting up this kind of backup is at risk of becoming outdated as AI tools evolve, but see 9 for a reference.

• A locally-installed LLM that doesn’t need a network connection and that someone else can’t change on you. See Nicolle Weeks’ guidance for beginners10.

• Backup copies of tool configurations that worked yesterday and last month.

• Clearly defined contexts that limit the risk of an updated model interpreting your instructions differently. See Hodman Murad’s guidance on this.

• A competent human who understands the work (note that the longer the work has been automated, the harder these people will likely be to find).

• A strategy on what tasks could best be postponed until AI service is restored.

Tactics like using a local LLM or gateway as a backup are contingencies that will reduce the impact if your main tool is disrupted. Using them as primary tools, not just as a backup, will make them mitigations that reduce the odds of your AI use being disrupted.

For additional mitigation, consider looking into ‘AI governance’ strategies and development tactics to help you identify specific risks beyond tool disruption. For instance:

• If you write code with AI, see Elena | AI Product Leader’s advice and templates for how to do rapid AI prototyping without technical debt.

• More generally, look at Iwette Rapoport’s guidance on using governance in the thread for human-AI collaboration.

• If you develop global products, check out Julia | Taking you global’s advice on AI, multi-lingual data governance, & content development.

The wisdom you need on how to identify and manage these AI dependency risks is out there waiting for you, in newsletters and in books and in chat groups. If you need help finding the resources you need, I invite you to reply, comment here, or drop me a DM.

Leave a comment

TL;DR Be prepared.

Keep using the AI tools, but build a risk management plan, and practice training the cognitive skills you want to keep.

And for heaven’s sake, as I’ve often said, please don’t shame anyone for using AI as a support tool. Let’s focus any shaming on AI tool companies who are behaving unethically; use & boost the few ethically-developed tools out there; and help folks make more-informed decisions on AI. I’ll be right there with you.

Leave a comment