Who Gets Credit for Ideas in Human AI Collaborations?

A few months ago, AI made headlines for cracking a mathematical puzzle that had stumped experts for decades.

Mathematician Paul Erdős posed the unit-distance conjecture predicting how many pairs of points could sit the same distance apart. AI proved it wrong by tracking down a group of infinite counterexamples to the solution Erdős proposed.

Although there was no single mathematician to congratulate, the solution was also the result of a kind of collaboration. After all, AI was trained on human data, and humans guided the search, even if it was hard to pin down who contributed what.

The Erdős problem is an extreme version of the attribution dilemma that Khoury College of Computer Sciences professor Christoph Riedl has been exploring in an effort to address the challenges that crop up as AI becomes part of everyday collaborative work.

“As people continue to collaborate with AI, both on an individual but also on a team level, it becomes increasingly unclear who is contributing to the work and who owns the work,” Riedl told Northeastern Global News. 

What happens when large language models (LLMs) have a much smaller hand in the final product? How do you acknowledge their use while making sure that human collaborators get credit for their contributions? And how do you improve the collaboration itself, making sure AI doesn’t iron out potentially important wrinkles in the argument.

That’s where it starts to get messy, Riedl said.

Together with Khoury College of Computer Sciences professor Saiph Savage, computer sciences Ph.D. student Kashif Imteyaz and other colleagues, Riedl published a project proposal that suggests a novel way to probe for answers — a workshop that will serve as a real-world testing sandbox for exploring and potentially solving problems that arise in human-AI collaboration. 

The event, which will take place in Salt Lake City, Utah, this October, will bring together a multidisciplinary group of researchers and practitioners to explore the dynamics of human-AI collaboration and brainstorm ways to fairly account for contributions, preserve a chain of responsibility and nurture creativity when teaming up with bots.

“Let’s just experience it to study it,” Riedl said.

And there’s plenty to experience and study. AI agents are increasingly baked into collaborative workflows. For example, the transcription software Otter.ai might turn a roomful of people’s comments into one neat meeting summary.

They’re still fixtures of individual work, too. Tools such as Jenni AI or SciSpace draft outlines and write explanations for computational results. Claude composes code. And there’s always ChatGPT for bouncing off ideas, polishing drafts or tracking down that molecule name that’s been on the tip of your tongue for the last 20 minutes.

It’s easy to miss how much editorial input bots have on content people write, share and eventually publish in scientific journals, Imteyaz said.

Institutions are trying to meet the moment. Back in 2023, the journal Nature called for scientists using bots as research assistants to disclose their use in methods or acknowledgment sections. Companies such as Wiley and Elsevier have made similar requests, sometimes asking researchers to include their prompts. And Anthropic recently announced that Claude models launched this month and later will include a watermark tagging text and images as AI-generated, although the method isn’t foolproof. 

For example, false positives are possible if the user’s language matches the bot’s. At the same time, the lack of a watermark won’t necessarily mean that AI had no hand in the manuscript.

In the absence of a watermark, voluntary disclosure might seem like a simple enough ask. Except AI use isn’t nearly that tidy. 

“You have some initial idea, you give it to ChatGPT. You co-think, co-iterate,” Imteyaz said. “It’s very hard to distinguish what is yours … and what is the whisper from the agent,” he added. And disclosures can’t capture contributions you never recognized as being AI-influenced in the first place, he explained. 

And asking researchers to include prompts simply doesn’t make sense, Riedl added. An interaction often amounts to a “multi-day conversation built on memories that the chatbot forms,” he said. The prompt only captures a sliver of the exchange and isn’t enough to recreate the response. 

With several people bringing bot-influenced ideas to the table, things get even murkier, Riedl said. Each person arrives without knowing exactly how much of an idea is their own. The group’s collective use of AI adds another layer of uncertainty. 

For example, while “cleaning up” meeting notes, AI might “restructure the argument in ways that redistribute credit for key ideas,” the researchers argue. They call the result “contribution dissolution” — a kind of authorship fog that can sabotage healthy collaboration from the get-go.

But credit isn’t the only thing at stake — it’s also the originality of the ideas themselves, Riedl said.

When synthesizing input from collaborators, AI has a tendency to paint everything with the same brush, reduce intellectual diversity and “smooth over frictions” instead of pushing back the way a human would, he explained. It’s a bit like accepting too many autocomplete suggestions: with AI at the helm, output starts sounding generic.

There’s also the flip side of attribution — responsibility.

AU governance expert Neda Maria Kaizumi told Northeastern Global News that while “AI can become an extraordinary ‘thinking partner’ for scientists,” it’s important to keep human reasoning firmly in the loop.

“If a researcher accepts an AI-generated hypothesis because it sounds convincing, who is responsible when it is wrong?” Savage asked. Preserving a “clear chain of accountability” can be difficult, she added.

Her answer is to keep people in the driver’s seat. “The more powerful the technology becomes, the more important human judgment becomes, not less,” she said.

The workshop will tackle all of these questions from the inside out.

Researchers will bring their unique ways of working and using AI to explore joint ownership and collaboration in real time and see what dynamics arise, Riedl said.

The workshop will also examine contribution dissolution and two related problems. The documentation trap asks why documenting AI use isn’t enough. Accountability infrastructure, in turn, explores what alternative systems might be needed to assign credit and responsibility without a paper trail clearly tracking contributions.

For example, in the “Documentation Trap Activity,” participants will receive a finished report, earlier drafts and a full log of the team’s AI interactions. They will then have to piece together who came up with what, whose ideas survived and who should be held accountable if something went wrong. 

Ultimately, the workshop is setting the stage for a paradigm shift around what it means to be an author in the first place. As AI takes on more of the execution, people increasingly focus on directing the creative process, Savage noted. 

“The author isn’t the hand anymore. The author is the mind that frames the problem,” she said, adding that it’s also “not a zero-sum threat to human creativity” but rather an extension of it.

Riedl has already seen hints of this creative potential in his research on human-AI synergy. When working on brain teasers, teams consisting of bots and people become more than the sum of their parts. 

“So they’re truly producing something that they couldn’t have produced individually,” he said.