What’s in a name? A comparison of attitudes towards artificial intelligence (AI) versus augmented human intelligence (AHI)

In 2018, the American Medical Association released a policy statement on augmented intelligence in medicine [1]. The wording “augmented intelligence” was carefully chosen in contradistinction to the more colloquial term “artificial intelligence” (AI) to emphasize that while computing systems have the capability to augment human medical decision making, these systems are not a replacement for rational human thought.

Popular culture and science fiction are plagued with examples of AI as a competing intelligence to be feared, and a recent survey of attitudes about AI among the general public found that only a minority support the development of AI [2]. In fact, most Americans believed that automation and AI would result in a net destruction of jobs [2]. Perhaps in response to these public perceptions, major cloud computing vendors, including IBM [3] and Microsoft [4] also have stated a preference to refer to the technology as “augmented” rather than “artificial” intelligence. Although the idea of “augmented human intelligence” (AHI) has been around for over 50 years, [5] the term “artificial intelligence” continues to be the prevailing term that is used. At our own institution, Mayo Clinic has established “augmented human intelligence” (AHI) as the preferred term, [6] yet our experience has been that in practice the two terms are often used interchangeably.

Perceptions of technology and attitudes toward these technologies are key elements that can affect the uptake and success of implementation. However, there is a lack of evidence as to whether there is a measurable difference in perceptions and attitudes toward the technology when it is referred to as “augmented human intelligence” versus “artificial intelligence” among health care staff. Therefore, in this study we sought to understand whether use of the less frequently used—but institutionally-preferred—term “augmented human intelligence” led to more favorable staff perceptions and attitudes about the technology.

Thirty-seven participants of the diabetes-related AI/AHI pilot and 56 of the hospital readmission-related AI/AHI pilot completed the survey (Table 1). The response rate was 46% for staff involved with diabetes-related AI/AHI and 38% for staff involved with hospital readmission-related AI/AHI, yielding an overall response rate of 41% (Table 1).

Survey responses are shown numerically in Table 2 and graphically in Fig. 1. Mean response score differences ranged from 0.04 to 0.22 points out of 5. No statistically significant difference was observed when comparing responses between respondents who were asked about AI versus AHI within each group (i.e., diabetes-related vs. hospital readmission related) or when combining the two groups (i.e., p > 0.05). At least 60% of respondents in each group did not think that AI or AHI would make their jobs obsolete. Respondents were largely ambivalent about the ability of AHI to understand their jobs.

At our institution, survey responses about attitudes following use of augmented human intelligence did not appear to vary when respondents were asked about AI or AHI. We were reassured by this, as we were worried that staff may misunderstand questions if asked using the institutionally-preferred term “AHI” rather than the more common term “AI.”

The finding that survey responses did not vary based on wording also suggests that staff perceived these two terms similarly. Although we did not directly assess perceptions of the two terms (e.g., by asking “How are AI and AHI different?”), this finding suggests that use of the term AHI does not necessarily “soften” attitudes or yield more favorable attitudes. This is similar to the findings of a survey of attitudes toward AI conducted in the general population [2], which compared attitudes toward “AI” and “machine learning” and found similar results for both terms.

One of the possible limitations of any negative study is whether it was powered to detect a difference that would be practically significant. Based on our sample size, we had a power of 95% to detect a difference greater than 0.8 points in our 5-point Likert scale, and an 80% power to detect a difference greater than 0.65 points. Although there is a possibility that a smaller difference existed but was not detected, we deemed differences less than one point to likely be of little practical significance.

However, we may have failed to detect variation in different attitudinal dimensions, and our study was not sufficiently powered to conduct sub-group analyses to determine whether attitudes differed between clinical work roles. We also only assessed the attitudes of staff at a regional health system clinic where a specific AI/AHI-based pilot project was being implemented, leaving the possibility that attitudes may vary in other settings where AI/AHI was not being implemented. In an initial survey, the common term “AI” was used. We cannot rule out the possibility that use of this term led to anchoring and contributed to the lack of an observed difference in responses between the groups. However, our observation that approximately 40% of respondents in reach group did not report being familiar with AI/AHI suggests that this may not have been the case for some staff. Indeed, it is possible that lack of familiarity and deep understanding of the terms (AI/AHI) accounts for the observation that there was no difference in responses when the two terms were used. During the project, we avoided using the term “AHI.” When this term was introduced in the survey presented in this manuscript, it did not seem to soften staff attitudes toward the technology, as some have hypothesized it would [3, 4, 6]. Attitudes of clinical staff elsewhere and of the general population (i.e., patients) may differ. Finally, because we targeted a convenience sample of staff participating in a quality improvement project, we must consider the degree to which selection bias may have influenced participant responses.

Our finding that attitudes did not appear to be different when technology was referred to as “AI” versus “AHI” can yield different interpretations. One interpretation is that the terms are perceived equivalently and can both be used interchangeably. An alternative interpretation is that, because the choice of term does not meaningfully influence staff’s perceptions about the work domain, other considerations should dictate which term is preferred. Our institutional leadership has chosen “augmented human intelligence” (“AHI”) as the institutionally-preferred term, rather than “artificial intelligence” (“AI”). To maintain consistency, we will utilize the institutionally-preferred in future pilot projects with clinical staff.

Although findings may be setting-specific, we observed that use of the terms “AI” and “AHI” in a staff survey on attitudes towards machine learning-based clinical decision support elicited similar responses.