As scientists and researchers seek to integrate AI more deeply into daily lab workflows, tackling the issue of false or misleading information is a top priority, says Andrew Wyatt, Chief Growth Officer, Sapio Sciences.
AI is increasingly important for biopharma R&D, offering breakthroughs in drug discovery, protein modeling, and disease prediction.
Tools like Insilico Medicine’s generative models and DeepMind’s AlphaFold have captured headlines and inspired optimism. That said, despite these advances, trust in AI tools remains fragile.
One of the most critical issues delaying adoption is AI hallucination – when models generate false or misleading information while presenting it as fact.
As researchers seek to integrate AI more deeply into daily lab workflows, addressing these hallucinations and the broader data integrity issues they stem from is becoming a top priority.
Several key factors will shape whether AI becomes a truly reliable scientific partner or remains a promising but risky frontier.
1. Building trust with quality data
For wider integration in drug discovery, AI must be trained using data that is both high quality and scientifically relevant. Today’s AI tools often rely on publicly available datasets, which may be outdated, fragmented, or lacking the contextual depth required for experimental work.
Proprietary datasets, real-time device readings, and unpublished research are frequently out of reach.
This lack of comprehensive data not only limits accuracy but contributes directly to hallucinations and misleading outputs. In fact, a recent Elsevier survey showed that 91% of researchers want LLMs trained exclusively on trusted sources.
Moving forward, the focus will be on building AI systems that can absorb and interpret complete, real-world lab data across formats, including structured results, lab notebooks, and instrument logs.
2. Addressing hallucinations as a fundamental design challenge
AI hallucinations are more than just a glitch, they reflect deeper weaknesses in how models are trained, fine-tuned, and deployed.
Hallucinations occur when a model fills in gaps in its knowledge with content that sounds plausible but isn’t grounded in evidence.
In scientific environments, this can lead to flawed hypotheses, wasted resources, and compromised research integrity.
The growing consensus among scientists and researchers is that AI tools must be held to higher standards of factual accuracy and source traceability. Future advancements will depend on designing AI models that include built-in mechanisms for fact-checking, uncertainty estimation, and contextual validation, ensuring outputs are not only fluent but also reliable and verifiable.
This is important because while hallucinations will reduce as the LLM’s access more data, it’s very difficult to regain user trust after even one flawed or inaccurate result.
3. Integrating real-time and in-lab data into AI workflows
Despite significant improvements in data capture, much of the information generated in modern labs remains siloed and inaccessible to AI tools.
Experimental results, metadata from lab instruments, and real-time observations are often scattered across disparate LIMS and ELN systems. Without access to this high-value, context-rich data, AI tools are limited to analysing static or outdated information, reducing their ability to adapt or improve over time.
The next phase of AI integration must involve connecting models directly to live lab environments, enabling real-time learning from experimental data and dynamic feedback from instruments. This shift is essential to create AI systems that reflect the evolving nature of scientific inquiry rather than static snapshots.
4. Improving usability and accessibility of AI in lab settings
Crucially, wider AI adoption in biopharma also depends on making tools more accessible to non-specialist users.
Historically, lab informatics systems have suffered from unintuitive interfaces, form-heavy workflows, and poor integration between platforms. AI tools – especially those powered by natural language processing and generative capabilities – offer an opportunity to simplify user interactions dramatically.
Emerging solutions allow scientists to describe experiments in simple language, build workflows through voice interfaces, and receive AI-generated suggestions in real time.
These innovations reduce the learning curve and empower more scientists to incorporate AI into their research. As virtual “co-scientists” become more capable, they’ll support everything from experiment planning to real-time decision-making without requiring users to become data experts.
5. From isolated successes to scalable, everyday tools
While the success stories in AI-driven drug discovery are compelling, they often remain exceptions rather than the rule.
Projects like AlphaFold2 or BioGPT demonstrate the technology’s potential but are typically led by well-resourced teams and supported by specialised infrastructure. For many labs, AI remains abstract – powerful in theory but challenging to implement in day-to-day research.
Future progress will be defined by how well AI can scale from high-profile experiments to routine lab work. This will require modular, interoperable systems that fit within existing workflows, as well as broader training and support to ensure scientists can trust and effectively use these tools.
AI continues to redefine what’s possible in biopharma, but its integration into research environments depends on overcoming persistent trust and usability challenges. As we look at 2026, success will hinge on the industry’s ability to eliminate hallucinations, unify data, simplify access, and align AI systems with scientific rigour.
Organisations that meet these challenges head-on will not only avoid the pitfalls of flawed outputs, they will lead the next era of faster, smarter, and more effective drug discovery.
