Healing Trauma!

Unlock the secrets to healing trauma! Explore how emotional coherence and metacognition can help you process difficult experiences and transform your identity. Learn how long to feel emotions and when to move forward.

#traumahealing #emotionalhealing #personalgrowth #metacognition #mindfulness #mentalhealth #selfhelp #NeuroFeedback #Neuropathy #NeuroSurgeon #NeuroDiversidade #NeuroPsychology #NeuroSciences #Neurologie #NeuroCoaching #NeuroRehab #wellbeing #innerpeace

Chatbots vs inline automation and their respective implications for students

In Generative AI for Academics I drew the distinction between conversational agents and templated responses. In the former you interact in natural language with a chatbot through a dialogue, whereas in the latter you select buttons to enact pre-defined transformations to text. This is how I talked about it in a webinar a couple of years ago:

https://youtu.be/GTjd9eLd9cw?si=bGt1KVk2CXKVpQv1&t=2859

Kim et al draw a similar distinction: “Two interaction techniques currently dominate: dialogue (e.g., OpenAI’s ChatGPT and Google’s Gemini) and predictive text completion (e.g., GitHub Copilot)”. I’d suggest that inline automation as found in Microsoft Copilot 365 or Gemini in Google Docs is a development of predictive text completion, in the sense that a user can complete a transformation of text by pushing a button. In both cases the system produces content with the difference being how this production happens: with chatbots it happens through dialogue and with inline automation it happens through pressing a button. If you’re a confident writer who uses chatbots in a purely dialogical way, it’s easy to forget how much chatbot use can be geared towards the production of content. I’d say it’s better educationally in the sense that some articulation is necessary to produce anything from a chatbot. But it’s still outsourcing a task to the machine in a radical way.

The approach developed by Kim et al tries for a third approach which “motivates users to reflect on their text with LLM-generated summaries, questions, and advice on writing (which we refer to as LLM views), helping them discover opportunities for improvement or elaboration”. They offer a useful definition of two stages involved in this process (my emphasis):

Revision means critically examining and evaluating, which we refer to as reflection, and identifying any opportunities for improvement or further development, which we refer to as discovery, and then making the appropriate changes. This can occur at any stage of the writing journey

The purpose of their interface is to offer perspectives on the text which scaffold reflection and discovery while avoiding the language model merely producing text directly for the user. They identified three outgrowths in their pilot of the project which are educationally relevant: (1) discovering underdeveloped ideas, (2) catering to their audience, and (3) identifying opportunities to improve clarity.

If we’re going to be using inline automation with students, this suggests how we can do it in a pedagogically responsible way. The purpose is to offer perspectives, taking advantage of how the model is embedded in the office software, rather than predictive text completion. Copilot offers these perspectives which present themselves for educational use. The problem is that (1) they are integrated seamlessly into an interface which is built around inline automation (2) they are not fine tuned for the specific domain so their advice can often be questionable.

But they are nonetheless there, which offers pedagogical opportunities. The key I think is to encourage students to gravitate towards dialogue (using the prompt windows where they are available) and perspectives (using the reflection tools built in) while discouraging them from using the push-button forms of automation. Where these are used, they need to be engaged with thoughtfully: to reflect on why you are selecting this button and to review the consequences of pressing it. The challenge is introducing moments of friction and reflection into software which is fundamentally designed to be frictionless, but it’s not insurmountable I think. Engaging with it educationally in this way though presupposes a lot of critical AI literacy in general and familiarity with Copilot 365 in particular.

Does RLHF wreck a language model's calibration? Not quite. The calibrated signal relocates rather than vanishes: a confidence the model states in words tracks accuracy better than its own token probabilities, often halving ECE. For instruction-tuned models, calibration becomes an elicitation problem more than a recalibration one. Just ask for the number.

https://benjaminhan.net/posts/20260610-just-ask-calibration/?utm_source=mastodon&utm_medium=social

#LLMs #Calibration #Metacognition #EMNLP #AI

What are the conditions which make it possible to learn with AI?

This thoughtful pre-print by Favero et al offered a pleasingly straight forward answer to this question. Ultimately we know what makes for effective learning:

• It has to involve active engagement in which the learner is reflecting, connecting and synthesising material rather than simply passively receiving it.
• It has to integrate the new understanding which emerges into their existing understanding through direct experience and dialogue.
• It has to involve the phased withdrawal of support as learners gain competence and confidence.

Rather than ask ‘what are the conditions which make it possible to learn with AI’ we can instead ask ‘what are the conditions which make it possible to use AI in these ways‘? The problem with commercial chatbots is not the technology itself but rather the design decisions involved in making a technology for a mass audience (as Nick Srnicek has plausibly argued AGI is ultimately an ambition to create a product which work universally without sector-specific fine tuning) which mean that, not only is it not adapted to these specific educational requirements, it actively works against them in a number of ways:

• The role of prompting has become decreasingly important with successive models
• The models are increasingly agentive in the sense of able and willing to actively do stuff for the user
• The role of model memory makes it easier for habits to form over time so that users get locked into certain ways of using the chatbot

This means there’s a fundamental tension in using these chatbots for educational purposes. It doesn’t mean it’s impossible. We also shouldn’t look the perfect be the enemy of the good: as the WonkHE report plausibly argued their widespread use reflects weaknesses of existing provision in that students are drawing on them to meet needs we are failing to meet. But we need to be realistic about the underlying tension, even as we try and mitigate it through AI literacy. In part this means helping students relate to chatbots in a way which doesn’t avoid difficulty. I really like how the authors describe the problem here:

However, learners often try to avoid such an effort. Studies show that high perceived difficulty and low short-term performance can discourage engagement, despite clear long- term benefits. Interestingly, Deslauriers et al. [29] found that students in active learning environments learned more but felt they learned less. Mental effort was misinterpreted as failure, while smooth lectures mistakenly felt more effective, though they were not. This cognitive bias leads students to favor fluent, low-effort activities that give the illusion of learning, such as re-reading polished explanations, without fostering deep processing [27, 30, 16].

AI tools, and particularly chatbots, may exacerbate this issue. By offering quick, fluent, and simplified answers, they reduce the cognitive struggle which is essential to learning [14]. Their convenience may lead to passive consumption, decreased research and reasoning skills, and a growing dependence on pre-digested knowledge [31]. Ease of use is appealing, but true learning comes from effort, complexity, and time.

This isn’t just significant for their studies. There’s AI slop proliferating in academic workplaces which appear to embody the same tendency, in which people in a rush can produce something superficially plausible which lets them tick it off a list and leaves them feeling more accomplished. Knowing how to sit with difficulty, to avoid the temptation of superficially fluent outputs, will I’m fairly confident be a skill that employers will value ever more in the coming years. This means recognising the difference between ‘desirable difficulties’ and contingent barriers which can be automated away. It’s impossible to do this unless use remains active throughout what Milan Sturmer and I call the user-model interaction cycle:

• Positioning: establishing a role for the model in interaction
• Articulation: putting ideas into words which are provided by the model
• Attunement: feeling the model has ‘understood’ what the user has brought to the interaction

Each of these dimensions can be (relatively) passive. Each of these dimensions can be active. The thread uniting active use is metacognition, in the sense the authors talk about here:

Routine use of AI tools can hinder the development of metacognitive skills, independent thinking, and intellectual agency [5, 12]. As students begin to outsource decision-making to the ma- chine, they risk becoming passive recipients of information rather than critical participants in the learning process [4]. This passivity is linked to broader deficits, including reduced creativity, increased mental laziness, and diminished capacity for critical thought [16]. Moreover, dependency on AI tools can lead to the uncritical acceptance of their outputs. When students perceive these systems as convenient, accurate, and reliable, they may stop questioning the information provided, which fosters cognitive dependency, i.e., the erosion of the ability to assess, verify, and challenge content independently

We need to help students name what it feels like to be actively thinking when interacting with a model. There’s a distinct phenomenology to this. It’s also something which cannot be sustained indefinitely. As you get tired, it’s easy to slide into increasingly passive uses of the model, with ever more capable models almost seamlessly picking up the load from you.

Generative AI and metacognitive laziness

While I’m sceptical of their experiment research design*, the concept of metacognitive laziness from this paper is clearly a useful contribution to thel literature. As Fan et al define it, this refers to “earners’ dependence on AI assistance, offloading meta – cognitive load and less effectively associating responsible metacognitive processes with learning tasks”. This matters because “offloading metacognitive effort to AI tools results in less effective engagement with essential self-regulatory tasks,” (pg 506). The risk is not just the offloading itself, it is increased passivity in the wider process of which the offloaded tasks are part.

This can undermine self-regulated learning because the metacognitive requirements for doing this effectively (e.g. goal setting, self-monitoring, self-evaluative etc) can be eroded over time by a reliance on the AI to negotiate difficulty. As they summarise the risk on pg 492:

the tendency of learners to become over-reliant on AI poses challenges for hybrid intelligence. This issue aligns with the concept of cognitive offloading, as proposed by Risko and Gilbert (2016), where learners delegate cognitive tasks to external tools to reduce cognitive effort. Although cognitive offloading can be beneficial in managing cognitive load, it may lead to decreased internal cognitive engage- ment over time, ultimately impacting learners’ ability to self-regulate and critically engage with learning material (Risko & Gilbert, 2016). Such cognitive offloading can lead to habitual avoidance of deliberate cognitive effort, a phenomenon echoing the emergence of what we term metacognitive laziness. From a more theoretical perspective, Alter et al. (2007) demonstrated that metacognitive experiences of difficulty or disfluency activate more analytical reasoning processes. When learners encounter situations that challenge their intuition, they are more likely to engage in deliberate analytical thinking (i.e., System 2 processes) (Alter et al., 2007). In the context of GenAI, if learners rely excessively on AI-generated outputs or facilitation, they might not experience the necessary disfluency or cognitive difficulty to trigger these deeper metacognitive processes.

The experience of difficulty activates metacognition. If the students cognitively outsource in increasingly habitual ways, it doesn’t just mean they lose the learning involved in what they are outsourcing. It means they lose their capacity to tolerate difficulty, as well to respond metacognitively to that difficulty. This points to the assumption which many educators have that there is something fundamentally corrosive in how students relate to AI which carries a threat exceeding the particular risks for any one assignment. This is a really sharp conceptualisation of the epistemic risk for learning involved in generative AI which gets beyond some of the limits of the ‘cognitive offloading’ concept.

*It seems fundamentally implausible to operationalise intrinsic motivation in the context of an experimental study. If you reduce motivation into the student’s expressed engagement with discrete tasks then it’s been quite dramatically circumscribed to fit the experimental constructs. Furthermore, we urgently need longitudinal studies in order to make meaningful claims about things like ‘cognitive off-loading’, ‘skill atrophy’ and ‘metacognitive laziness’. These just aren’t things which can be studied adequately at the level of discrete tasks, particularly ones that have been designed by a research team and have no real stakes for participants.

Pluralistic: Refining humanity (05 Jun 2026)

https://fed.brid.gy/r/https://pluralistic.net/2026/06/05/defining-humanity/

Sometimes, I feel like I spend too much time in metacognition 🤪

#psychology #neuroscience #metacognition

https://www.psychologytoday.com/us/blog/brain-reboot/202511/why-does-your-brain-hate-learning?ref=refind

#metacognition concept map

#thinking

https://www.patreon.com/posts/159869522

10 posts about #metacognition -- a vital skill in the age of social media and #ai

#thinking

https://www.patreon.com/collection/2124752?view=expanded

Teaching #metacognition techniques to 4-6 year olds leads to better learning outcomes (and may immunize against cognitive decline caused by #ai use)

#llm #learning #education

https://www.patreon.com/posts/159662223