Note: this article was originally written in Dutch, commissioned by the Dutch magazine for journalists: Villamedia
Dear Laurens,
I have only just got used to the idea of chatbots. And now I am already hearing about so-called AI agents. Are these autonomous assistants something I as a journalist should be scared of or dream of? Help me! – Emma
Dear Emma,
“You know what you should have?” a friend said, two years ago, after my book came out and I was getting a lot of requests for speaking engagements thanks to the introduction of ChatGPT. “An agent!” I decided to take his advice. For a small fee, I now outsource almost all my negotiations, client communications, agenda management and invoicing to Kiki and get feedback, advice and unsolicited tips in return. She knows my work and senses where and with which audience my story resonates. She thinks along and keeps contacts warm. I grant everyone this agent, but there is only one Kiki, unfortunately. And that is what AI software developers want to change.
What is an AI agent?
It is essentially a digital assistant that, at your behest, acts independently online. It differs from ‘regular’ AI chatbots (such as ChatGPT) in three things. Agents:
1. search the current web,
2. reason logically, and
3. take action independently, outside the chatbot interface.
The creators of chatbot Claude.ai recently showed their agent ‘Computer Use’, in which two AI agents work together to operate a computer and design a 1990s website and get it working in a browser. Another example: if you want to get the phone numbers of all Danish libraries neatly put in a spreadsheet, soon all you have to do is give an agent the task and you will receive an Excel file with the data in your mail not much later.
Of course, Emma, there are already chatbots that search the web and claim to be able to reason. For instance, a new model for ChatGPT uses something called ‘chain of thought’: the model makes a kind of thought leap, comes up with a plan and executes it.
Salesforce and Microsoft are therefore positioning their agent technology as the next logical step after AI chatbots: it will not only speak to companies’ customers, but also immediately arrange the returning of shipped items, for example. Providers of AI agents promise efficiency, scalability and 24/7 availability.
What will journalists be using AI agents for?
AI agents can take over routine journalistic tasks, such as searching for certain people on Linkedin or summarising posts on 100 different Bluesky accounts.
Journalism researchers Joris Veerbeek and Nick Diakopoulos went a step further. They worked out a scenario with three AI agents taking on the roles of analyst, editor and reporter. In doing so, the first processes large amounts of data, the second questions its outcomes and instructs the third agent to investigate further and search some more.
Tool-based and simulation-based agents
The above examples are in the new realm of so-called ‘agentic AI’ of the ’tool-based’ type. They autonomously fill spreadsheets, send emails on your behalf and schedule appointments. They save time, are practically supportive and, above all, production-oriented.
Where things get even more exciting (or scary) is when tech companies lead us to believe that these new AI agents are human in their thinking and reasoning. That sounds good, but as Volkskrant journalist Laurens Verhagen aptly noted, “it’s mostly a piece of brilliant marketing“. James O’Donnell of MIT Technology Review sees a danger when tool-based agents merge with those ostensibly human agents we call simulation-based. To that type you command: act like a journalist, biologist or teacher. According to O’Donnell, the two will eventually merge into a new kind of AI software, which not only mimics your personality or identity, but also goes out digitally to act on your behalf (or indeed, as you).
How this can also play out in the offline world is illustrated both hilariously and frighteningly by investigative journalist Evan Ratliffe in his podcast The Shell Game. In it, you follow the AI clone of Evan’s voice who, linked to ChatGPT, makes independent phone calls to his wife, helpdesks and scammers.
Why you can’t trust these agents
Casey Newton, tech-journalist and founder of Platformer warns of the dangers and risks that Ratliffe also demonstrates in his podcast: AI agents do not always act as we want or expect. Who is liable if your AI agent makes mistakes? What happens to the data it collects? Microsoft’s early AI agent Recall took screenshots of everything you did on your computer… not a good idea.
Important criteria for journalists’ use of AI agents are about transparency and accountability: how do you guarantee quality and reliability when AI agents do the work on your behalf? Maybe it’s okay if you have an agent pre-calling 10 experts to check if they can join your talk show tonight. But on what basis does such an agent then decide whether the expert fits the context and is competent?
“Which model would you trust to act on your behalf?” asks Mairéad Pratschke, professor of Digital Education at the University of Manchester. She points to the recently demonstrated ideological premises of different large language models: ‘When we move from the stage of simply generating output to creating agents who perform tasks for us, but without necessarily checking in when they make decisions, it’s more than a little frightening.’
Expand AI guidelines for agents
Agents, like chatbots, seem impressively capable of performing delineated tasks. Although they fairly autonomously figure things out for you, they lack the judgement and moral compass that human journalists (hopefully) do possess. You can’t ask an AI agent to ‘sense’ something or assess newsworthiness. Our finesse and intuition simply cannot be externalised to software-based systems.
So Emma, some concern is again in order here. Not because AI agents are necessarily bad, but because we need to define what type of agent we want to use in our journalistic process. And where and what we may want to outsource.
Because… let me ask you this: who exactly are you helping when an AI agent generates a summary of your video item on your behalf and automatically posts it on BlueSky? Let AI agents do some tasks just fine, but let us journalists judge the quality and value of our work above all.
Why we still need humans – maybe more than ever?
Sometimes Kiki calls me to ask how I am doing or to tell me what the audience thought of my lecture. She also invites me to New Year drinks and asks if I have any great weekend road tips for getting away in the Eastern part of the Netherlands. I never asked her for this, but they are exactly why I value my human agent so much.
Be surprised by people, not technology!
–
Build your own AI agent in 20 minutes (as recommended to me by the wonderful Nele Goutier:
Need more visionary AI agent ideas? Then check out this video from AI browser builder Arc: https://www.youtube.com/watch?v=C25g53PC5QQ&t=342s