DARPA posts AI-related RFI
The Defense Advanced Research Projects Agency (DARPA) posted a request for information (RFI) on Identifying Artificial Intelligence (IAI). Responses are due by 4:00 p.m. Eastern on May 1.
DARPA’s Defense Sciences Office (DSO) is requesting information on foundational concepts and potential technologies to identify a system that is “artificially intelligent” in a way that would qualitatively change how we interact with such a system. While there is no consensus definition of what constitutes “Artificial Intelligence” (AI), for the purposes of this RFI, AI refers to an engineered embedded capacity of a system to perform complex tasks and / or adaptively learn. AI capabilities can cover a vast space with varying roles, functions, sophistication, and autonomy. Knowing a priori where in this space the AI is operating in a system can substantially change how humans would interact with such a system. As such, we are interested in the question of how one can passively or actively determine that there is intelligence embedded in a system and what capabilities it enables.
Responses should address the following elements:
1. Identification of a particular system (e.g. vehicles) or application domain (e.g. game playing) where AI-based methods are currently (or soon will be) creating significant changes in capability.
2. For the identified system, a definition of the capability classes1 within the system/domain chosen that you believe would lead to qualitatively different behavior, and methods that might be used to infer the capability classes (as you have defined them) from these signals.
3. The “avenues of exploitation” for identifying the presence of AI? Consider at least these four types of access: a. “White box” – internal system access up to (but not including) source code. b. “Black Box” – external signals including inputs and outputs as well as side channel signals such as EM, thermal, or acoustic emissions. c. “Standoff” – passive monitoring of activity under normal operation at maximum physical range d. “Interact” – perturbing the operational environment (e.g. introducing innocuous inputs to test the response) or the AI itself, external to what it is embedded in.
4. A discussion on the following: a. Other types of access that might exist; b. Available signals at each access level for detecting capability; c. Information revealed about the system’s capability at each access level; d. The timescales needed to collect such information, and what might the computational or other resource costs be to process the information.
5. Finally, suppose the capability classes were not fully known in advance. In that case, discuss what methods might be used to infer what might be the full set of capabilities from the signals described above. Respondents who wish to address more than one system/domain should answer the questions above for each (in its own section or via a separate submission)
Note that this RFI is not interested in any Turing test, alternative Turing test, AI “IQ tests,” or botnet detection methodologies.
Full information is available here.