Reducing echo chamber effects: an allostatic regulator for recommendation algorithms

Objective

Recommendation systems are prevalent on the Internet but are prone to feedback loops that cause “echo chamber” effects. These effects can have negative consequences for users’ well-being, diversity of information, and social cohesion. Therefore, there is a need for novel techniques to combat echo chamber effects and promote healthier online experiences.

Method

We present an allostatic regulator for recommendation systems based on opponent process theory. This regulator can be applied to the output layer of any existing recommendation algorithm to dynamically restrict the proportion of potentially harmful or polarised content recommended to users, based on the users’ recent content history. We implement our prototype algorithm as a code wrapper for a supervised K-Nearest Neighbors algorithm for movie recommendations and evaluate its performance using simulated user data.

Results

Our results show that allostatic regulation is effective at reducing echo chamber effects in a simulated population. The method can be used for regulating the entire range of possible online content and can adapt to evolving user behaviours.

Conclusions

The allostatic regulator is a promising technique for mitigating echo chamber effects, providing app developers with a flexible tool to help users self-regulate their online experiences.

2.1. Allostatic regulation model

Our allostatic regulator was mathematically defined using a behavioural dose-response model based on the concept of “intervention dose”, in which dose-response effects scale upward with increasing exposure to an intervention or behaviour (Heerman et al., Citation2017; Henry et al., Citation2023; Kirk et al., Citation2019; Manojlovich & Sidani, Citation2008; McVay et al., Citation2019; Voils et al., Citation2012). This model characterises a human behaviour through four main attributes: $\mathit{Potency}$ (level of engagement in a behaviour, with arbitrary units), $\mathit{Amount}$ (time spent on the behaviour, in seconds), $\mathit{Frequency}$ (how often the behaviour occurs in a given time, in Hz), and $\mathit{Duration}$ (the observation period, in seconds). Then, the dose for a single behaviour can be calculated as:

(1) $\mathit{Dos}{e}_{\mathit{individual}\mathit{behavior}}=\mathit{Potency}\times \mathit{Amount}$(1)

and the total dose of cumulative behaviours performed within a specific period is:

(2) $\mathit{Dos}{e}_{\mathit{total}}=\mathit{Frequency}\times \mathit{Duration}\times \overline{\mathit{Dos}{\mathit{e}}_{\mathit{individualbehaviors}}}$(2)

where $\overline{\mathit{Dos}{\mathit{e}}_{\mathit{individualbehaviors}}}$ represents the mean individual behavioural dose for all behaviours over the $\mathit{Duration}$ in which $\mathit{Dos}{e}_{\mathit{total}}$ is assessed. To simplify our analysis, we performed a Behavioral Frequency Response Analysis (BFRA) by treating $\mathit{Potency}$, $\mathit{Amount}$ and $\mathit{Duration}$ as constants and varying the $\mathit{Frequency}$ at which the dose is delivered (Henry et al., Citation2023).

We assumed that the subject’s response to the behavioural dose took the form of opponent processes with exponential decay, a widely used framework in neuroscience and behavioural science (e.g., see Koob & Le Moal, Citation2008; Pirih et al., Citation2022; Solomon & Corbit, Citation1974; Uribe-Bahamonde et al., Citation2019). In this model, the subject’s hedonic response to the behavioural dose consisted of exponentially decaying a- and b-processes with decay constants ${\lambda }_a$ and ${\lambda }_b$, with units seconds⁻¹, and the $\mathit{Potency}$ of the behaviour was modelled as the initial intensity of the process, in hedons. We chose to represent the a- and b-processes as Hawkes processes with exponential decay kinetics, obeying the laws of superposition. This parallels the strategy employed by researchers studying the simulation of social network dynamics, in which user activity is represented using stochastic point process models, often with an exponential decay component (Dai et al., Citation2017; Du et al., Citation2015; Farajtabar et al., Citation2018; Pan et al., Citation2016; Tan et al., Citation2016; Trivedi et al., Citation2017; Wang et al., Citation2016, Citation2017, Citation2018).

Given a behaviour (in this case, watching a movie) performed at times $\mathit{T}=\left\{t_1,\dots ,t_{\left|\mathit{T}\right|}\right\}$, where $\left|\mathit{T}\right|$ is the total number of repetitions of the behaviour, we can define the hedonic state $\mathit{h}\left(\mathit{t}\right)$ of the subject at time $\mathit{t}$ by convolving $\mathit{f}\left(\mathit{t}\right)$, consisting of Kronecker delta functions denoting the times when the behaviour occurred and its potency, with $\mathit{g}\left(\mathit{t}\right)$, the sum of the a- and b-processes:

(3) $\mathit{f}\left(\mathit{t}\right)=\mathit{P}\sum _{\mathit{i}=1}^{\left|\mathit{T}\right|}\mathit{\delta }\left(\mathit{t}-t_i\right)$(3)

(4) $\mathit{a}\left(\mathit{t}\right)=I_a{e}^{-{\mathit{\lambda }}_{\mathrm{a}}\mathrm{t}}\mathit{t}\ge 0$(4)

(5) $\mathit{b}\left(\mathit{t}\right)=-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\mathrm{t}},\mathit{t}\ge 0$(5)

(6) $\mathit{g}\left(\mathit{t}\right)=\mathit{a}\left(\mathit{t}\right)+\mathit{b}\left(\mathit{t}\right)=I_a{e}^{-{\mathit{\lambda }}_{\mathrm{a}}\mathrm{t}}-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\mathrm{t}}$(6)

(7) $\mathit{h}\left(\mathit{t}\right)=\mathit{f}\left(\mathit{t}\right)\ast \mathit{g}\left(\mathit{t}\right)$(7)

which can be condensed to the general formula:

(8) $\mathit{h}\left(\mathit{t}\right)=\mathit{P}\sum _{t_i\le \mathit{t}}^{\left|\mathit{T}\right|}\left(I_a{e}^{-{\mathit{\lambda }}_{\mathrm{a}}\left(\mathrm{t}-{\mathrm{t}}_{\mathrm{i}}\right)}-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\left(\mathrm{t}-{\mathrm{t}}_{\mathrm{i}}\right)}\right)$(8)

where $\mathit{t}$ is time; $\mathit{P}$ is the $\mathit{Potency}$ of the behaviour (hedons); $\mathit{\delta }\left(\mathit{t}-t_i\right)$ is the delta function at time $t_i$ (seconds); $I_a$ and $I_b$ are the initial intensities of the a- and b-processes (arb. units); and ${\lambda }_a$ and ${\lambda }_b$ are the decay constants of the a- and b-processes (seconds⁻¹). By setting $\mathit{Duration}=t_{\mathit{sim}}$, and by fixing $\mathit{Amount}$ and $\mathit{Frequency}$ as constants, we can observe the effect of modifying the values of $\left|\mathit{T}\right|$, $\mathit{Potency}$, ${\lambda }_a$ and ${\lambda }_b$ on the user’s predicted hedonic state $\mathit{h}\left(\mathit{t}\right)$ over time.

2.2. Applying allostatic regulation to a movie recommender

To apply allostatic regulation in a real-world setting, we trained a movie recommender on the MovieLens-1 M dataset, containing 6,000 user ratings on 4,000 movies, and their genres (Harper & Konstan, Citation2016). We wrote a program in Python 3.8 (Van Rossum & Fred, Citation2009) to simulate the movie-watching behaviours of virtual subjects. In this hypothetical scenario, each subject employed allostatic regulation to reduce the proportion of movies recommended with the “Horror” genre, to lower the amount of anxiety-provoking material that they were exposed to (this was an arbitrary choice; as another example, the subject could have chosen to regulate “Romance” or “War” movies to reduce exposure to sexual or violent content). It was assumed that these subjects enjoyed watching Horror movies, but only infrequently, as too many Horror movies can lead to excessive anxiety and difficulty sleeping (Martin, Citation2019). Thus, the aim was to regulate the frequency of Horror movies watched to keep the user within their hormetic limit.Footnote¹ The code and figures for the simulation can be found in the Supplementary Materials.

The proportion of movies of a single genre watched by all virtual subjects during the simulation time $t_{\mathit{sim}}$ can be described by the equation:

(9) ${\mathrm{\Omega }}_j=\frac{{\sum }_{i=1}^n{\sum }_{j=1}^x{m}_{\mathit{ij}}}{\mathit{Total}}$(9)

where ${\mathrm{\Omega }}_j$ is the proportion of movies of genre $\mathit{j}$ watched by all virtual subjects over $t_{\mathit{sim}}$, $\mathit{n}$ is the number of subjects, $\mathit{x}$ is the total number of genres, $m_{\mathit{ij}}$ is the number of movies of genre $\mathit{j}$ watched by subject $\mathit{i}$ during $t_{\mathit{sim}}$, and $\mathit{Total}$ represents the total number of movies watched by all subjects over $t_{\mathit{sim}}$.

The initial movie watched was “From Dusk Till Dawn (1996)” which contained the “Action”, “Comedy”, “Crime”, “Horror” and “Thriller” genres. Each virtual subject watched one movie every 72 hours, randomly selected from a list of the 20 top recommendations generated by a K-Nearest Neighbors (KNN) algorithm from the Scikit-learn package (Pedregosa et al., Citation2011). The KNN algorithm, a supervised machine learning method, assigned labels to new data points based on the labels of their closest neighbours in the training data. It identified these neighbours by calculating distances between data points and selecting the K nearest ones. The new data point was then assigned the label most common among its neighbours. In this scenario, the algorithm recommended movies similar to the most previously watched movie only, using genres and mean ratings as the features for training the model. Mean ratings were calculated for all movies with at least 10 ratings, with the remainder of movies excluded from the dataset. Previously watched movies were excluded from the latest recommendations. Although the recommender’s optimisation was modest, it effectively served our objectives.

Allostatic regulation was applied by modifying the KNN algorithm using the “AdjustedKNN” class, which can be found in the second Supplementary Materials file. Specifically, the Euclidean distances for Horror movies were dynamically increased in proportion to the inverse of the user’s hedonic score, $\mathit{h}\left(\mathit{t}\right)$. Hence, if the user watched multiple Horror movies in quick succession, the regulator’s allostatic load would increase rapidly, causing the proportion of recommended Horror movies to decrease in response. For this model, a-processes were excluded, meaning that the value of $\mathit{h}\left(\mathit{t}\right)$ was a pure model of allostatic load for b-processes only. Parameter exploration involved testing different b-process parameters within plausible ranges $\left(\mathit{I}\_\mathit{b}=\left\{0.08-0.16\right\},\mathit{\lambda }\_\mathit{b}=\left\{0.0035-0.0069s^{-1}\right\}\right)$, with higher values of $I_b$ and lower values of ${\lambda }_b$ leading to stronger regulation.

All code and figures for our simulations can be found in the Supplementary Materials.

3.1. Regulating the entire possible range of content with b-processes

In opponent process theory, the b-process persists longer than the a-process. Therefore, when behaviours occur at time intervals shorter than the critical decay duration (CDD) of the b-process (defined here as the time it takes for the b-process to decay to insignificance) but longer than the CDD of the a-process, then only the b-process contributes to allostasis. Under such conditions, it is unnecessary to model the a-process to simulate allostatic effects. Thus, to achieve basic regulation, we removed the a-process from the allostatic model and treated $\mathit{h}\left(\mathit{t}\right)$ as a scalar weighting factor that modified the recommendation probability of different types of content. Then, the opponent process model $\mathit{g}\left(\mathit{t}\right)$ simplifies to:

(10) $\mathit{g}\left(\mathit{t}\right)=\mathit{b}\left(\mathit{t}\right)=-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\mathrm{t}},\mathit{t}\ge 0$(10)

And $\mathit{h}\left(\mathit{t}\right)$ more closely models the allostatic load generated by the b-processes (i.e. how far the hedonic state deviates from homoeostatic levels, on average):

(11) $\mathit{h}\left(\mathit{t}\right)=\mathit{P}\sum _{t_i\le \mathit{t}}^{\left|\mathit{T}\right|}\left(-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\left(\mathrm{t}-{\mathrm{t}}_{\mathrm{i}}\right)}\right)$(11)

demonstrates how exponential b-processes can take a wide range of shapes which, when combined with an allostatic paradigm, allows us to regulate any type of content. One practical example is the regulation of online hate speech. This is a contentious issue due to the need to balance freedom of expression with prevention of harm and illegal activities (Banks, Citation2010). Here, we demonstrate how one may perform allostatic regulation of the entire spectrum of possible online discourse, ranging from positive to hate speech. Let us consider these two extremes.

First, we take the case of online hate speech, which in certain contexts and jurisdictions, is illegal (Alkiviadou, Citation2019). Given a behavioural dose represented as a delta function $\mathit{\delta }\left(\mathit{t}\right)$ delivered at $\mathit{t}=0$ seconds with $\mathit{Potency}=1$ hedons, and convolving it with a b-process $\mathit{b}\left(\mathit{t}\right)=-I_b{e}^{-{\lambda }_b\mathit{t}}$ with initial intensity $I_b$, time $\mathit{t}$ in the interval $\left[0,\mathrm{\infty }\right)$ and decay constant ${\lambda }_b=0$ yields the person’s hedonic state over time:

(12) $\mathit{h}\left(\mathit{t}\right)=\mathit{P\delta }\left(\mathit{t}\right)\ast -I_b{e}^{-0\mathit{t}}$(12)

$=-I_b\mathit{\theta }\left(\mathit{t}\right)$

where $\mathit{\theta }\left(\mathit{t}\right)$ is the Heaviside step function:

(13) $\mathit{\theta }\left(\mathit{t}\right)=\left\{\begin{array}{c}1,\mathit{t}\ge 0\\ 0,\mathit{t}\le 0\end{array}\right.$(13)

This results in a step function that does not decay. Hence, using ${\lambda }_b=0$ seconds⁻¹ in combination with a large value of $\mathit{P}$ (for example, $\mathit{P}=\mathrm{\infty }$) creates a large step function that could be used to regulate – or even permanently block – harmful content.

On the other end of the spectrum, harmless content, regardless of frequency, can be given a decay constant ${\lambda }_b=\mathrm{\infty }$. Then, convolving the delta function with $\mathit{g}\left(\mathit{t}\right)$ yields:

(14) $\mathit{h}\left(\mathit{t}\right)=\mathit{P\delta }\left(\mathit{t}\right)\ast -I_b{e}^{-\mathrm{\infty t}}=0$(14)

Thus, it is possible to classify and regulate the entire range of content – from positive speech to hate speech – by varying ${\lambda }_b$ and $I_b$ between $\mathrm{\infty }$ and 0. demonstrates this range of modulation. Setting ${\lambda }_b=\mathrm{\infty }$ seconds⁻¹ has no regulatory effect on the algorithm’s output as it causes $\mathit{g}\left(\mathit{t}\right)$ to immediately decay to zero. Such a function can be used for harmless positive content; meanwhile, setting ${\lambda }_b=0$ seconds⁻¹ generates a b-process that doesn’t decay, but grows each time the behaviour is performed. Hence, it is the value of ${\lambda }_b$ that is most important, as it defines the rate of allostasis.

The true utility of this method lies in adaptively regulating hormetic content. For instance, reading new articles online in small doses lets the reader stay informed and learn different perspectives. However, excessive news consumption can lead to addiction, doom-scrolling, and anxiety due to negative content (Anand et al., Citation2022; Shabahang et al., Citation2021). To counter this, one could use allostatic regulation to dynamically restrict their frequency of consuming news articles with negative content. shows how a b-process with decay constant ${\lambda }_b=0.0029$ seconds⁻¹ (a half-life of four hours) can reduce the number of negative articles recommended. If the person read multiple articles in a row (from $\mathit{t}=240$ seconds onwards), then rapid allostasis occurs, reducing the number of negative article recommendations. However, if only a single negative article is read in isolation ($\mathit{t}=40$ seconds), then the b-process is allowed to decay back to homoeostatic levels. Users should be allowed to modulate ${\lambda }_b$ to match their own hormetic limit of article consumption.

3.2. Simulating allostatic regulation in a movie recommender

shows the simulated proportion of movie genres watched by all subjects over time. Simulations with () and without () regulation are presented. Horror movies are presented in green in the centre of each Figure. Applying the regulator reduced the proportion of Horror movies recommended, and watched, during the simulation. Interestingly, even at high levels of regulation (), periods with no Horror movie recommendations were temporary. During these periods, the high allostatic load initially prevented any Horror movie recommendations, but then subsequently decayed to allow Horror movies to be recommended in moderation.

3.3. Quantifying the effects of different allostatic parameters

A comparison of () shows the effect of doubling $I_b$ had approximately the same regulatory effect as halving ${\lambda }_b$. This can be explained by calculating the integral of the b-process over the time interval $\left[0,\mathrm{\infty }\right)$, which is proportional to the contribution of the b-process to allostasis:

(15) ${\int }_0^{\mathrm{\infty }}-I_b{e}^{-{\mathit{\lambda }}_{\mathrm{b}}\mathrm{t}}\mathit{dt}=-\frac{I_b}{{\lambda }_b}$(15)

Therefore, $I_b$ and ${\lambda }_b$ are inversely proportional, in terms of their contribution to allostasis. Given that the half-life of the b-process, $t_{1/2_b}$ (seconds) is related to the decay constant by the equation

(16) $t_{1/2_b}=ln\left(2\right)/{\lambda }_b$(16)

then we can also conclude that $t_{1/2_b}$ is proportional to $I_b$; hence adjusting either variable will have similar effects on ${\mathrm{\Omega }}_{\mathit{Horror}}$, the proportion of Horror movies watched. However, the rate of allostasis initially will be higher for the b-process with the lowest value of $t_{1/2_b}$ (or the highest value of ${\lambda }_b$), as demonstrated in : while both b-processes have the same integral and the same steady-state, the initial rate of allostasis is highest for the b-process with the highest value of ${\lambda }_b$.

To test the hypothesis that adjusting $t_{1/2_b}$ would have similar effects to adjusting ${\lambda }_b$ over an extended time, we conducted 150 simulations by varying $I_b$ between 0 and 0.3, and 150 more simulations varying $t_{1/2_b}$ between 0 and 300 hours, to determine the relationship between these variables and ${\mathrm{\Omega }}_{\mathit{Horror}}$. demonstrates the relationship between $I_b$, $t_{1/2_b}$, and ${\mathrm{\Omega }}_{\mathit{Horror}}$, with trendlines derived via piecewise regression. To the right of the discontinuity in each Figure, the behaviours were performed at a time interval shorter than the CDD of the exponential, preventing the exponentials from decaying sufficiently before the next behaviour was performed. This led to increasing rates of allostasis, producing an exponential reduction in the percentage of recommended Horror movies as $I_b$ or $t_{1/2_b}$ increased.

However, an asymptotic limit can be seen in each Figure, because as frequency increased towards infinity, the rate of allostasis was balanced by the decay rate of newly formed b-processes. Hence, given an infinite simulation time and assuming ${\lambda }_b>0$ seconds⁻¹ (i.e. $t_{1/2_b}<\mathrm{\infty }$ seconds), the percentage of Horror movies watched can never fully decrease to 0, as an equilibrium will always form between the rate of allostasis and the generation rate of new b-processes. This equilibrium represents a balance where the user is neither overexposed to nor completely deprived of this type of content.

3.4. Allowing behavioral bursts by incorporating a-processes

Our movie regulator was conservative and only included b-processes, as demonstrated by the dark blue curves in . However, if the time interval between behaviours is shorter than the CDD of the a-process, positive allostasis (sensitisation) can occur before negative allostasis (habituation). This is demonstrated in the green opponent process curves in , representing combined a- and b-processes. At a behavioural frequency of 0.1 minutes⁻¹ (), the CDD of the a-process is shorter than the time period between behaviours, and so any positive allostasis effects are negligible. However, at a behavioural frequency of 0.5 minutes⁻¹ (), the CDD of the a-process is longer than the behavioural period, leading to an initial rise in $\mathit{h}\left(\mathit{t}\right)$ for the combined opponent processes (transient sensitisation). Eventually, b-process allostasis exceeds that of the a-process, leading to negative allostasis.

Including a-processes in the allostatic model allows for more sophisticated regulation of different behavioural patterns, such as short behavioural bursts that may be beneficial in certain contexts. For instance, a person may wish to study a topic by watching several videos in quick succession. In this case, it may be helpful for the algorithm to suggest videos on the topic for an hour, before introducing some regulation to prompt a break from learning. A similar method could be applied to allow brief periods where news articles or social media posts can be accessed in moderation during work breaks. illustrates these “behavioral bursts”, simulating a person who consumes multiple pieces of digital content in short periods of time. In this case, regulation was applied to end the behaviour as soon as the a-process no longer exceeded the person’s baseline hedonic state. While this shows the versatility of allostatic regulation, more research is needed to verify its effectiveness in such scenarios.

4.1. Summary of findings

In this article, we have introduced a prototype allostatic regulator which can be applied to recommendation systems as a code wrapper – an additional layer of code that adds functionality to the recommendation system without modifying the original code. This technique provides a dynamic, anticipatory solution for reducing echo chambers that form on social media, and can be tailored to the individual based on their preferences for regulation. While our results focused on the regulation of a movie recommender, allostatic regulation can in theory be used for regulating the outputs of any type of recommendation system using the principles of opponent process theory. For example, it could be used to limit exposure to extreme political content and conspiracy theories, reduce the frequency of sensationalised news articles, and prevent overexposure to addictive, pornographic, or violent material on social media platforms.

Empirical validation of this technique should involve controlled user studies, in which participants interact with recommendation systems with and without allostatic regulation, measuring engagement, content diversity, and psychological outcomes. Furthermore, A/B testing on live platforms could assess real-world effectiveness by analysing metrics such as watch time, click-through rates, and user retention. Economic analyses could also be conducted to determine whether implementing allostatic regulation aligns with the financial interests of social media companies, demonstrating feasibility and improving industry adoption.

4.2. Theoretical implications

Allostatic regulation may help to limit polarisation and echo chamber effects by moderating the extremes of content recommended to users. It achieves this in two ways. Firstly, the regulator dynamically adjusts the allostatic load based on the recency and frequency of behaviours with exponentially decaying b-processes. If the allostatic load is not allowed to decay back to homoeostatic levels, then the regulator will naturally adjust the set point of the system, such that a new steady state is reached; this prevents positive feedback loops (Goldstein & McEwen, Citation2002). Secondly, as shown in , the regulator never eliminates Horror movies completely, but always allows some non-zero proportion of Horror movies to be recommended. The equilibrium observed between decaying b-processes and the rate of allostasis means that content diversity is preserved without completely suppressing user preferences. In terms of a solution for political content, an allostatic regulator can ensure that users are occasionally exposed to differing viewpoints, even if most of the content aligns with their preferences. Our proposed approach may empower users and ensure they are not trapped in echo chambers for specific content categories. “Random nudges” could also expose the user to diverse views to reduce polarisation (Currin et al., Citation2022). However, this may not be sufficient; Noordeh et al. (Citation2020) found that prolonged exposure to collaborative filtering recommendations on the MovieLens-1 M dataset reduced content diversity, leading to echo chambers that individual users could not escape by simply adjusting their own ratings. This suggests that further interventions may be required to prevent echo chamber formation.

To employ allostatic regulation, app developers must find the optimal level of regulation that produces beneficial effects while not being too restrictive (Brown, Citation2020). Users may reject the regulator if it is perceived as biased or overly censorious. It may be desirable to employ a user-centred approach to regulation, where app developers include simplified virtual dials that allow users to adjust their content diversity without needing to understand allostasis. For example, the developer could group potentially harmful content into different categories (such as pornographic, violent, or political content), and apply allostatic regulation to each category. Users could then adjust the dial for each category to fine-tune the amount of regulation applied to each type of content.

The allostatic model of hormesis is flexible enough to classify “grey area” content, which can be defined as content consumed in small doses or frequencies up to a hormetic limit, beyond which the content becomes harmful. An example of such content is conspiracy theories, which may have some value in stimulating critical thinking and exposing possible abuses of power. However, consuming many conspiracy theories may increase experiences of anxiety, uncertainty aversion, and existential threat (Liekefett et al., Citation2023). Further, people have different levels of receptivity to conspiracy theories, which their previous exposure may influence (Douglas et al., Citation2019). Each person may have a unique hormetic limit for conspiracy theory content, influenced by their environment, genetics, and upbringing. As such, individuals may require that their allostatic parameters be uniquely quantified to keep them within their hormetic limits.

To perform this quantification, longitudinal methods like Ecological Momentary Assessment (EMA) may prove valuable, in which participants’ emotional states are tracked over time and measured at regular intervals (Shiffman & Stone, Citation1998). The parameters for the a- and b-processes specific to individuals consuming certain types of content could be modelled from this EMA data once it becomes available. In the meantime, we can make some general assumptions about the shape of these opponent processes based on the current literature. For example, we can assume that stronger stimuli lead to larger opponent processes, based on studies such as the one by Roberts and David (Citation2023), who showed that smartphone users experience a greater flow state and higher degrees of subsequent depression when using the TikTok app compared to other mobile apps.

4.3. Limitations and future directions

One limitation of this study is that we did not explore the effects of allostatic regulation across multiple genres. Our focus on the Horror genre was arbitrary and served only as a demonstration; similar regulatory effects would likely emerge for any genre. However, future research should examine how allostatic regulation operates when applied to multiple genres simultaneously, with each genre having unique decay and potency parameters. This would better reflect real-world situations, where users can consume content with mixed genres in parallel and may need to regulate their use of multiple genres at once. Indeed, this will be one of the primary challenges in integrating such a system into real-world recommendation algorithms: finding ways to both accurately classify and regulate content with mixed genres such that users can have their content automatically regulated.

Yet even if these technical challenges are overcome, a fundamental dilemma remains: what incentive do major social media companies have to adopt such a system? Since allostatic regulation reduces the frequency of highly engaging but potentially harmful content, companies may argue that it risks decreasing user engagement and, consequently, advertising revenue. However, in the long run, platforms that integrate responsible content regulation may benefit financially by fostering healthier user behaviours, reducing burnout, and improving public perception. By positioning themselves as ethical alternatives to unregulated competitors, these platforms may attract users who prioritise digital well-being. If voluntary adoption proves unlikely, top-down regulatory intervention may become necessary, though government-imposed solutions risk being perceived as heavy-handed and could face resistance from both industry and users. A balanced approach – where platforms are encouraged to self-regulate while being held accountable for algorithmic harm – may be the most viable path forward.

Allostatic regulation can be seen as a form of adversarial training, in which the machine learning algorithm makes predictions, and the allostatic regulator provides an alternative objective function for the algorithm to optimise. We used a simple form of supervised learning – the KNN algorithm – to demonstrate that it is possible to adjust the model weights even after training. Importantly, this method is potentially scalable to more complex algorithms, including neural networks. This implementation may allow for regulation of more complex machine learning behaviours beyond recommender systems; for example, it may provide a method of classifying and regulating behaviours in wider branches of AI.

Subsequent research should explore the integration of this mechanism within the training process. The simplest approach involves altering the final layer of a machine learning algorithm or neural network. A more advanced method could involve adjusting the internal functions of the algorithm using the hedonic state as a multiplicative factor. For example, consider a recurrent neural network with Long Short-Term Memory (LSTM) architecture (Hochreiter & Schmidhuber, Citation1997). By using allostatic regulation to shift the sigmoid function’s centre in the forget gate of an LSTM network, one could regulate the “forgetfulness” of the network concerning different types of content. This could provide greater dynamic control of the network’s outputs than can be achieved by modifying the final network layer. Still, there is reason to avoid modifying the inner workings of algorithms. For example, a recent study revealed a decline in OpenAI’s GPT-4 model accuracy over time (Chen et al., Citation2023). Indeed, OpenAI’s technical report on their GPT-4 model (Achiam et al., Citation2024) demonstrates that Reinforcement Learning with Human Feedback (RLHF) improves model performance on some metrics but worsens it on others. This highlights an advantage of implementing allostatic regulation as a code wrapper around the machine learning model, which does not directly affect the model’s internal functions. Instead, it allows the model to learn user preferences while acting as a dynamic filter for the recommender’s outputs. Future research should compare the strengths and weaknesses of both methods.

In conclusion, we have created a prototype allostatic regulator that can provide dynamic content regulation for social media users, improving content diversity and reducing harmful content exposure. This regulator can theoretically be applied at the output layer of even the most complex recommendation algorithms, and represents a novel paradigm shift in digital content regulation, offering a psychology-based solution that can help minimise the formation of echo chambers online. While this technique works in a simulated environment, more empirical research is needed to quantify its effectiveness in real-world situations.