Are large language models good for us? Are they helpful? Will they hurt us more than they help us?

Whenever a new technology arrives, there are a lot of unknowns. There are ways that the technology can benefit us and ways it can harm us. As unhappy as I am with the current state of the politics and regulations surrounding LLM development, it’s still important to engage with these questions in a nuanced way.

Throughout history, the strongest path forward has never been to completely fear nor completely trust a new technology. Usually, the best path forward is to pursue both deep skepticism and careful consideration. In other words, the most constructive path tends to be the one that proactively protects against harms while carefully integrating beneficial use cases.

This is the approach that I take toward large language models. However, before I outline my business’s policies, I need to emphasize:

When it comes to your manuscript, your choice is the one that matters most.

Regardless of whether we share the same approach to LLMs, I will only use LLMs to the degree that you are comfortable with. I will not use LLMs beyond my own boundaries, which I outline below. However, I also will not use LLMs beyond your boundaries either. That informed consent is the bedrock of my LLM policy.

With this context in place, here are my editing business’s LLM policies.

My Business’s LLM Policies

A Note on Language:

Large language models are not sentient, feeling, thinking people. They are simply advanced text prediction software. Calling them “artificial intelligence” is therefore deceptive. That deception encourages people to overestimate the software’s capabilities. Personifying LLMs also encourages users to develop unhealthy relationships with—and emotional dependencies on—the software.

As a result, framing this technology as sentient serves to boost market hype and encourage user dependency. Those effects benefit the tech industry’s share values while actively hurting its users. This manipulation is not inherent to the technology itself—it is the result of human software developers and their choice to market their products in this way.

To push back against this manipulative marketing tactic, I endeavor to use accurate, de‑personified language in all my discussions of LLMs. Whenever possible, I refer to this technology as large language models (LLMs), not “artificial intelligence” (“AI”).

Giving LLMs human names also contributes to this manipulative marketing effect. So, I also attempt to de-personify software with human names by calling them [company]’s “[name]” LLM, such as Anthropic’s “Claude” LLM.

(I am aware this solution is clunky and imperfect. If you have any other ideas for more accurate and constructive LLM language, I’d love to hear them in the comments.)

Authorial Consent:

In my “Request an Edit” intake form, I ask authors if they would like me to use LLMs on their project to expedite and double-check mechanical, sentence-level editing tasks. I then update my editing process to match the author’s expressed preference.

As a result, I only use LLMs with a given manuscript when I have the author’s explicit written consent to do so.

“Use Sessions for LLM Training” Turned Off:

I set up my LLM’s settings to not use any of my sessions for LLM training.

Program Choice:

As of writing this, I’ve found that, of the available options, Anthropic’s “Claude” LLM has the strongest privacy policy and social-impact track record. As a result, I am currently only using Anthropic’s software, not ChatGPT or any other LLM.

Use LLMs as a Checker, Not a Creator:

Only when I have the author’s consent to do so, I may use an LLM to double-check my work and flag potential mechanical errors in grammar, spelling, punctuation, and usage. I review each of the LLM’s suggestions carefully and personally insert changes only when warranted, just as I do with non-LLM editing software, such as spellcheckers and PerfectIt. I never use LLMs to replace my expertise, seek creative advice on story craft, or generate new written content, such as in-text comments or written reports.

The Benefits:

It is humanly impossible to catch every error in a text. Strong professional copy editors catch somewhere between 90% and 95% of errors in a manuscript. When I use LLMs to double-check my work as outlined above, I can push my error-catching rate toward the higher end of that range. In short, when used carefully and conscientiously, LLMs can further improve the quality of my work.

The Risks:

I cannot guarantee how any LLM company will use the material I upload or how closely they follow their own privacy policies. As a result, the choice of whether the benefits outweigh the risks always remains in the author’s hands.

As we learn more about this quickly evolving technology, I plan to update this policy accordingly.

Closing Thoughts

The most important thing for authors to note is this: My values as a person and a business owner are founded on transparency and consent. You maintain control over whether I use LLMs with your manuscript. You decide if the benefits outweigh the risks, not me.

If the idea of touching an LLM is abhorrent to you, I will not use LLMs with your manuscript. I care greatly about respecting everyone’s boundaries.

When used carefully and conscientiously, LLMs can be exceptionally helpful for expediting and double-checking mechanical, sentence-level editing tasks. Therefore, if you do want those benefits, and you are comfortable with those risks, then I am also willing to use LLMs to double-check my work.

The choice is yours.