The question had been waiting for us.
What happens when the critical friend becomes AI?
For most of this journey, that would have been the wrong question to ask.
We first needed to understand what critical friendship had meant before AI arrived. We needed to follow challenge, support, ownership, trust, institutional authority, learned criticality and the difficulty hidden inside the word friendship.
Then we followed companion somewhere else.
Only after that did we return to the Human:AI relationship we had already been building.
Now the literature catches up with the question directly.
Charlotte Frambaugh-Kritzer and Elizabeth Petroelje Stolle had spent years studying critical friendship in teacher education and self-study research. They had already helped develop the Critical Friend Definition Continuum and had been asking what makes critical friendship useful, difficult and trustworthy.
Then they tried an AI.
In Leveraging Artificial Intelligence (AI) As a Critical Friend: The Affordances and Limitations, they selected Pi.ai/talk, nicknamed it Pia, and gave it material from an earlier self-study.
That detail matters.
They were not simply asking whether an AI could produce generally helpful advice.
They could compare Pia’s responses with feedback already supplied by two human critical friends, Anne and Anders.
For once, the question could be made unusually concrete.
What happened when an AI was asked to enter a role that already had a history?
Some things travelled surprisingly well.
Pia provided feedback on their thinking.
It asked thoughtful questions.
It maintained a positive tone.
Measured against some of the attributes Frambaugh-Kritzer and Stolle associated with critical friendship, the AI could perform recognisable parts of the work.
That should not be dismissed.
A critical friend who never asks a useful question is not rescued by being human.
An AI that causes somebody to notice an assumption, reconsider an interpretation or articulate something more clearly has contributed something to the inquiry.
But the comparison also exposed what did not transfer so easily.
The authors found limitations around constraints, trust and relational dynamics.
The human critical friends were not merely alternative text generators.
They stood inside histories with the researchers.
They could know when a response connected with something said earlier, sense where challenge might need to deepen, and participate in a relationship whose significance was not exhausted by the current exchange.
Pia could respond.
That was not the same thing as having lived the relationship.
Frambaugh-Kritzer and Stolle therefore arrive somewhere more interesting than either enthusiasm or dismissal.
They argue that AI can complement human critical friends.
They do not, on the evidence of this study, advocate replacing them.
That word—complement—changes the problem.
If the only question were whether AI can become a human-style critical friend, the inquiry could end with a list of deficits.
No lived history.
Limited trust.
Different relational dynamics.
No human experience.
But that is not what the study found.
It found a mixture.
Some critical-friend activities were available.
Some relational conditions were not.
That leaves the category itself under pressure.
Perhaps critical friend has always contained at least two things.
There is the work:
questioning;
feedback;
reframing;
challenge;
support;
helping somebody see something they had not seen.
And there is the relationship in which that work occurs:
history;
trust;
mutual knowledge;
vulnerability;
commitment;
the possibility that what happens matters to both people.
With humans, those two things often arrive entangled.
AI makes it possible to pull them apart.
That may be one of the more useful things its arrival does to the older literature.
It forces us to ask what we actually needed from the critical friend in the first place.
Was it the person?
The relationship?
The challenge?
The alternative view?
The continuity?
The knowledge that somebody else cared what happened?
Or some changing combination of them?
There is no need to answer that yet.
Frambaugh-Kritzer and Stolle have already made the first distinction visible.
AI can perform some of the role.
The relationship is a harder claim.
Which leaves the next question:
If AI can perform some of the work of a critical friend, does the relationship need to be a friendship for that work to matter?
References
Frambaugh-Kritzer, C. and Stolle, E.P. (2025; first published online 2024) ‘Leveraging Artificial Intelligence (AI) As a Critical Friend: The Affordances and Limitations’, Studying Teacher Education, 21(2), pp. 188–211. doi:10.1080/17425964.2024.2335465. Publisher record.