Academic Safety Papers – Talk but no Walk
A main part of my current research is investigating how we can assure the safety of decision-making in Autonomous Systems. Naturally this requires a lot of trawling through academic papers to see what already exists ‘out there’ – which can be summed up by the immortal words of a diminutive magician from yester-year…“not a lot”. Not outside of the AAIP anyway.
Faced with a torrent of ‘relevant’ returns, we are all well-versed in the process of eliminating ‘spurious returns’ – which is something along the lines of:
- Check the source (reputable Journal?)
- Check the authors (grey literature is rarely of use in my experience to date)
- Read the abstract (and aim of the paper)
- If the abstract looks good, read the conclusion
- If the conclusion confirms the author(s) established the aim, then collate with all the papers that are stored ‘for reading’.
…and then the disappointment and frustration creeps in.
It staggers me how many academic papers (from reputable outlets) have all the talk and no substance. I find all too often that extant papers that have been marked as relevant (using the above criteria from searches relevant to assuring the safety of decision-making in AS):
- Have ‘safety’ (or ‘safe’) in the titles but don’t discuss what safe actually constitutes
- Have ‘safety’ (or ‘safe’) in the titles but discuss only reliability
- Have ‘safety’ (or ‘safe’) in their titles, but never use the word again out of the abstract – although they do mentions accidents (mainly collisions) (but not the hazards that may realise them).
I know that I am fortunate to be working in a field that has limited answers currently – not least as there is a huge opportunity to keep me meaningfully employed! But with finite time available each week, it isn’t too much to ask for accurate indicators of what is in a paper…we’re not ready for academic click-bait are we?
The rest of the papers we (Richard Hawkins and I) are currently reading as part of our research seem to either take long-established safety engineering processes and re-purpose them with new titles (the process remains identical when you look beyond the new acronyms), or start from the point that safety requirements are already elicited, or are predicated on an assumption that what constitutes safe behaviour is already known.
Our research (which we’ll tell you all about VERY shortly) starts from the root of the issue – here’s a bit of a teaser from a planned paper:
“Decision making is fundamental to what constitutes an AS … This independent and autonomous decision making could give rise to new hazards and/or causes not present in an equivalent human-controlled system, yet the research into this area is lacking…We compare and contrast the differing casual models of autonomous and non-autonomous systems, and build on extant safety engineering techniques in order to assess and assure the process of autonomous decision-making; specifically considering the contribution of the operating domain”
Watch this space…