AI is not an "it"

We talk a lot in journalism about whether we should or shouldn’t use “AI.” Whether “AI” is good or bad. Whether the use of “AI” should be disclosed, prohibited or governed in certain cases.

But AI is not an “it.”

Treating it as a monolith is not only unproductive; I’d argue it weakens our ability to assess and adopt future technologies with the nuance those decisions require.

Is the recommendation system that powers the “recommended reads” module on your homepage AI? Do we need to disclose that to readers?

Is your predictive paywall AI? Does it feel more ethical, less ethical or simply different from other forms of audience targeting?

Is the NLP classifier that helped the Los Angeles Times report this story AI? Should a guild have a say in when machine learning and statistical inference are applied to reporting?

What about your CMS’ auto-tagging feature? Grammarly autocomplete? Your iPhone’s automated voicemail transcription? Even simple automations like web scrapers?

Should those things be governed by your organization’s AI policy, even though many of them have been in use for years? Should they be subject to AI restrictions in union contracts?

I have heard the term “AI” used to describe all of them.

Some edge cases are absolutely worth discussing. One minor consequence of generative AI’s arrival is that it has given more people a reason to pay attention to the current state of technology. More people are asking more questions, and there are real benefits to that.

But here’s my fear: Given the hostility toward AI in newsrooms right now, painting everything with the same broad brush risks moving us backward, or at least making us appear needlessly out of touch. It’s not hard to imagine an overbroad backlash against certain uses of generative AI spilling into debates over technologies and practices that have long been accepted.

As we lay down rules and norms around how these tools should and should not be used in our profession, it would benefit us to stop treating AI as a monolith.

We should distinguish between generative AI and automation, which may not involve AI at all. We should distinguish between machine learning, which has been used in newsrooms for years, and large language models, which are different in important ways.

The same is true even within the category of large language models. There is a meaningful difference between using an LLM to generate publishable text probabilistically and using that same LLM to extract information from source material in a way that can be checked, verified and constrained.

Not everyone needs to become an expert. But some shared literacy would help us ask better questions, write better policies and avoid wasting time fighting the wrong battles.

The question shouldn’t be whether and how journalism should use “AI.”

The question should be: Which technologies? For which purposes? Under what constraints? With what risks? And with what benefits?

That’s a harder conversation. But it is also a much more useful one.