With hallucinations waning, AI is expanding into scientific research



Scientific research has become one of the most promising applications for artificial intelligence. It’s also become one of the most divisive.

Across the world, tech firms are directing AI’s analytical power into tools for academics. Gradually, they’re entering every stage of the research process.

Today’s scientists can use TLDR to summarise study papers, Litmaps to find research gaps, Consensus to discover insights from expert scholars, and HeyScience for peer review.

Investors have taken a growing interest in the progress. Last year, Elicit research workflow system raised a cool $9mn within months of its launch, while fellow Californian startup NobleAI secured €17mn for a chemical and material formulation platform.

Across the Atlantic, European challengers are also emerging. The continent’s current leader is Iris, an Oslo-based scale-up that today announced a €7.6mn Series A funding round.

Iris’ flagship product is a machine learning engine that analyses academic research. Users prompt the tool to search for the information they need. The system then categorises, summarises, and systematises millions of documents to deliver actionable insights.

Alongside the academic users, corporate clients such as R&D firm Materiom have signed up as clients. So has the Finnish Food Authority (FFA), which used the platform to guide a strategy for bird flu.

“Our tools enabled them to rapidly comb through tens of thousands of research papers to pinpoint relevant information at the intersection of niche areas like farm biosecurity and migratory bird patterns — an analysis that would have taken researchers months to do manually,” Anita Schjøll Abildgaard, CEO and co-founder of Iris, told TNW via email.

Abildgaard is also bullish about the product’s accuracy. She claims to have a remedy for one of AI’s worst afflictions: hallucinations.

Fighting AI hallucinations

A major flashpoint erupted after Meta launched Galactica in November 2022. Billed as a large language model for science and trained on 48 million research documents, the software produced shortcuts to academic insights. Unfortunately, it also produced endless streams of AI-generated nonsense.

The results caused a social media uproar. Academics were also alarmed by AI’s propensity to generate inaccurate information. Researcher Michael Black, director of the prestigious Max Planck Institute for Intelligent Systems in Germany, warned the tool could “usher in an era of deep scientific fakes.”  

The reaction spooked Meta. Just three days after launching Galactica, the company pulled the plug on the tool.

As the controversy swirled, Iris was promoting a solution to these errors.

The startup blends knowledge graphs, extracted data, and context similarity testing to measure the factual accuracy of AI-generated content. To substantiate the results, the system also provides verifiable sources for the outputs.

With these safeguards in place, Iris promises to dramatically reduce hallucinations.

“By cross-checking the content generated by our AI against structured knowledge bases and evaluating the semantic similarities to factual sources, we can improve the factuality of our outputs,” Abildgaard said.

Armed with renewed cash supplies alongside the accuracy protections, Iris now plans to add extra tools to the arsenal. Up next is a chat assistant that will guide researchers through the system and personalises the workflow.

Yet accuracy remains the priority, Abildgaard said, because “factual grounding is paramount” in research.



Source link