Any automated scientific discovery system or Robot Scientist clearly has both advantages and disadvantages. Some of the main points are discussed here.
It has been suggested that systems such as ours would be better described as 'Laboratory Assistants' rather than the implicitly more independent term 'Scientists'. Although in some ways the term 'Laboratory Assistant' has some merit, as they are not independent workers. In other, more important ways, the term is inappropriate as Laboratory Assistants do not generally form hypotheses, decide on the experiments to test them, automatically analyse and interpret the results etc. It should also be remembered that these systems are still just prototypes, and it is probable that future developments in hardware and software will increase the independent nature of such systems. So, on balance, we prefer the more evocative term 'Robot Scientist', and argue that the Adam system has discovered new knowledge about gene function in S. cerevisiae
that has been independently verified [24
Another common argument against Robot Scientists is that they remove the chance for serendipitous discovery, and that they are incapable of innovation. We would argue that more often than not a serendipitous discovery is simply the result of an experiment that has been designed without prior analysis of all the potential outcomes. Louis Pasteur phrased this sentiment as 'In the fields of observation chance favours only the prepared mind'. While it is true that the underlying artificial intelligence components fail to meet human expectations for innovative thought, we believe that developing a richer background model and incorporating more sophisticated reasoning mechanisms will bring us closer to that goal.
It is also true that a Robot Scientist generates its hypotheses based on information obtained from publicly available databases, and as such is susceptible to any errors contained therein. However, this is no different to the situation in which human scientists find themselves, as they also have to rely on published information. Most such databases are curated by humans, and provide a service that biologists routinely use in their work. Both could choose (or in the case of the Robot Scientist be programmed) to assign weightings in their confidence of various pieces of evidence based on how they were labelled in the databases (e.g. indirect vs direct experimental evidence). When errors are present the robot is most likely to propose an incorrect hypothesis which the experimental data will then refute. For Adam we avoided problems by primarily using only one public database (KEGG) and manually updating our system model where conflicts were noticed, before allowing automated generation of hypotheses. Where a Robot Scientist may have more trouble is where it lacks the broader background knowledge base which may be available to a human scientist (e.g. the problem with Adam's system model not representing kinases mentioned earlier). A better system model and a broader knowledge base can be developed for Robot Scientists that in time would negate this difference.
Similarly, it has been pointed out that the data analysis algorithms of a Robot Scientist might be less able to deal with flaws in experimental measurements than a human, and may come to false conclusions as a consequence. We believe this to be mostly a matter of refinement of programming; for example, Adam's growth curve smoothing and de-noising routines use machine learning and statistical data analysis of multiple replicates to routinely deal with the effects of significant noise, contamination, and even gaps in the measured readings. Further refinement could be done to identify abnormal or other unexpected results, for example the shape of a bacterial contamination growth curve, or the long lag time associated with yeast cross-contamination, and then discount them automatically. The advantage of a Robot Scientist here is that it would always be consistent in its handling of the data. More physical issues with an experiment, for example flaws in plastic wares, faults in instruments, incorrect placements of plates in instruments etc. are currently easier for a human to notice and correct. Whilst there are some measures we can put in place to automatically deal with this type of issue (e.g. calibration and fault detection system checks prior to experiment runs), we believe that future refinements of the hardware and plastic wares used in laboratory automation will reduce the effects of this type of problem.
Finally, there has also been discussion about costs, comparing the cost of using a Robot Scientist against using humans to perform the same tasks. There is a substantial cost for these systems, not only in initial capital outlay and user training, but also in ongoing servicing and maintenance costs, and we would not currently consider them to be 'cost-effective' in comparison to human scientists. However, these systems are early prototypes, and we would expect such costs to reduce significantly as laboratory automation becomes more widespread, more reliable, and the software more user-friendly. The cost of hiring human scientists and technicians and buying the instruments and equipment they need to perform such high-throughput and complex experiments should not be underestimated either, and the robots have the advantages of efficiency, consistent quality, and the ability to run outside normal working hours.