Beyond the agreement paradigm: Regulating algorithmic collusion under the Competition Act, 2002

Antitrust laws presume a meeting of minds. Section 3(1) of the Competition Act, 2002 prohibits any agreement causing an appreciable adverse effect on competition, and Section 2(b) extends this to any “arrangement, understanding or action in concert.” Capacious as it is, that formulation did not anticipate pricing algorithms that, through reinforcement learning, collude autonomously without any human communication.

The concern is empirically grounded. Calvano, Calzolari, Denicolo and Pastorello showed in 2020 that independent Q-learning agents converge on supra-competitive prices and adopt reward-punishment strategies equivalent to cartelisation. Assad, Clark, Ershov and Xu found in 2024 that algorithmic pricing in German fuel retail raised margins by nine to twenty-eight per cent. The OECD’s 2017 analysis grouped these risks under design, structure and outcomes; Harrington proposed to attach liability to algorithmic behaviour, not agreement. Yet, none specifies the cumulative conditions, evidentiary standards or statutory form needed to operationalise liability.

This article fills that gap. It proposes a structural predisposition test (SPT), a three-part cumulative standard and a new Section 3A imposing liability for algorithmic coordination without disturbing Section 3 or the agreement paradigm.

The SPT is an independent basis for a rebuttable presumption of anti-competitive coordination. It neither redefines agreement nor stretches the existing framework, but opens a parallel pathway triggered only where three cumulative conditions converge.

Cumulation is central: it spares lawful oligopolistic interdependence while ensuring collusion-prone deployments do not escape merely because no agreement is provable.

The first condition: Algorithm design

The first condition requires architectural features showing a structural capacity for collusion: whether it can observe competitors’ pricing, directly or through shared feeds; whether it optimises on cumulative rather than single-period profits; whether it adapts to rivals’ reactions; and whether it lacks randomisation against deterministic convergence.

These features make autonomous collusion possible: the Calvano experiment showed that agents meeting them sustain supra-competitive prices. Algorithms that merely automate rule-based pricing, or optimise on internal demand without reference to rivals, do not qualify.

Assessment need not inspect proprietary code; it proceeds on firm-submitted documentation, within the CCI’s existing Section 36 powers.

The second condition: Market structure

The second condition requires deployment in a market conducive to collusion, marked by high concentration, transparent pricing, homogeneous products and frequent transactions, excluding algorithms that, though capable of convergence, operate where collusion is unlikely. The CMA’s 2018 working paper identifies these very factors as determinants of collusion risk.

Concentration may be operationalised through a Herfindahl-Hirschman Index threshold of 2,000, the benchmark for highly concentrated markets in United States and European Commission merger control. Indian merger review under Section 6 already relies on such analysis.

The third condition: Observable market effects

The third condition requires sustained effects consistent with collusion and inconsistent with competition: price elevation above competitive benchmarks, reduced dispersion, dampened volatility, coordinated responses to shocks and deterred entry. The observation period should run at least 6 months to exclude seasonal or shock-driven noise. Liability thus attaches only where demonstrable harm results, screening out the short-term parallelism common and lawful in oligopolies.

Once the conditions are met, the burden shifts to the deploying firm, which may rebut by showing that its algorithm operates independently of competitor signals, that the market lacks coordinating features, that the outcomes reflect cost shocks, demand fluctuations or differentiation, or that effective safeguards exist. This preserves the line between unlawful coordination and lawful parallelism.

The SPT cannot be accommodated within the existing Section 3 as currently drafted. Even the “action in concert” formulation of Section 2(b) presupposes conscious human coordination and redefining agreement to capture algorithmic convergence risks incoherence. A cleaner solution is a parallel statutory provision that preserves Section 3’s structural integrity for cases involving human coordination while creating an independent pathway for algorithmic conduct. The proposed Section 3A reads:

3A. Anti-competitive algorithmic coordination.

(1) No enterprise or association of enterprises or person or association of persons shall deploy or cause to be deployed any algorithmic pricing system in respect of production, supply, distribution, storage, acquisition or control of goods or provision of services, which causes or is likely to cause an appreciable adverse effect on competition within India, notwithstanding the absence of any agreement within the meaning of section 2(b).

(2) Where an enterprise or association of enterprises or person or association of persons engaged in identical or similar trade of goods or provision of services deploys an algorithmic pricing system and;

(a) Such system,

(i) Observes, collects or responds to information relating to the pricing or commercial conduct of competing enterprises, whether directly or through any shared data feed or third-party intermediary,

(ii) Determines or adjusts prices or other commercial terms with reference to long-run or cumulative returns through any adaptive, learning or automated process that conditions future pricing upon the observed or anticipated conduct of competing enterprises; and,

(iii) Does not incorporate such constraints or safeguards as would prevent the system from converging upon coordinated pricing outcomes;

(b) Such deployment is in a relevant market which, having regard to the level of concentration, the degree of pricing transparency, the homogeneity or substitutability of goods or services, and the frequency of transactions among competing enterprises, is conducive to coordinated outcomes; and

(c) Such deployment results in, or is likely to result in, sustained pricing or other commercial outcomes consistent with coordination and inconsistent with independent competitive behaviour, over such period as may be specified by the Commission, having regard to the factors specified in sub-section (3) of section 19, such deployment shall be presumed to cause, or to be likely to cause, an appreciable adverse effect on competition within India.

(3) The presumption under sub-section (2) may be rebutted where the enterprise establishes that,

(a) The algorithmic pricing system was independently designed and operated in a manner that does not satisfy the conditions specified in clause (a) of sub-section (2);

(b) The relevant market does not exhibit the characteristics specified in clause (b) of sub-section (2);

(c) The observed market outcomes are attributable to changes in costs, demand conditions, efficiencies or other lawful commercial considerations; or

(d) The algorithmic pricing system incorporates such constraints or safeguards, including randomisation mechanisms or limitations on the use of competitor pricing data, as materially reduce the likelihood of coordinated pricing outcomes.

(4) Any person or enterprise that designs, develops, supplies or operates an algorithmic pricing system which is deployed by two or more competing enterprises engaged in identical or similar trade of goods or provision of services shall be deemed to have contravened sub-section (1) where such person knew, or ought reasonably to have known, that such system, when so deployed, was capable of producing outcomes of the nature referred to in clause (c) of sub-section (2).

Provided that nothing contained in this sub-section shall apply where such person demonstrates that reasonable measures were taken to prevent the system from producing such outcomes.

(5) Nothing contained in this section shall restrict the deployment or use of any algorithmic pricing system which does not satisfy the conditions specified in sub-section (2), or which determines prices solely with reference to the costs, demand, inventory or other internal commercial considerations of the deploying enterprise without observing, collecting or responding to the pricing or commercial conduct of competing enterprises.

(6) The Commission may, for the purposes of any inquiry under this section, require any enterprise or person to furnish such information, documents or technical disclosures relating to the design, operation, objective function, data inputs or functioning of any algorithmic pricing system as the Commission may consider necessary.

Explanation: For the purposes of this section, “algorithmic pricing system” means any automated computational system that determines, recommends or adjusts prices or other commercial terms for goods or services through the processing of market data, competitor information, or any combination thereof, and includes any system employing machine learning, reinforcement learning, artificial intelligence or other adaptive computational methods for such purposes.

Any framework imposing liability must satisfy constitutional requirements, chiefly Article 14 (equality), Article 19(1)(g) (freedom of trade) and Article 19(6) (reasonable restrictions).

The classification satisfies Article 14: the distinction between algorithms meeting the three conditions and those that do not rests on clear, objective and verifiable criteria, and serves to curb consumer harm from coordinated pricing.

It also engages Article 19(1)(g): algorithmic pricing is commercial activity, so any restriction must satisfy Article 19(6). The provision is narrowly tailored, since liability arises only where collusive design, conducive structure and demonstrable effects coexist. Sub-section (5) preserves freedom to deploy algorithms outside them.

Even where the conditions coincide, the presumption remains rebuttable. The framework does not prohibit algorithmic pricing; it adapts competition law to reach what agreement-based enforcement cannot.

Ex-post enforcement alone is insufficient. Adoption is expanding and harm may accrue long before detection, so the framework adds phased ex-ante governance.

First, within 18 months of enactment, the CCI should establish a Digital Markets and Algorithmic Analysis Unit staffed in data science and computational economics, beginning with market studies in high-risk sectors such as e-commerce, travel and fuel retail.

Second, the CCI would administer mandatory audits in markets designated by the second condition, requiring disclosure, in confidence, of the objective function, data inputs, reward structure, optimisation horizon and safeguards, feasible under existing Section 36 powers.

Third, compliance-by-design obligations would impose safeguards at the design stage: limits on conditioning prices on competitor signals, mandatory randomisation against deterministic convergence and constraints on retaliatory capacity. Sandboxing would let firms test high-risk systems under CCI supervision, with safe harbour for cleared systems.

Third-party liability under sub-section (4) ensures that developers, vendors and platform intermediaries supplying systems to competing firms bear a proportionate share of responsibility. The RealPage litigation in the United States illustrates the point: a shared platform producing coordinated rent increases exposes not only the landlords but the platform itself.

Coordination with sector regulators is essential: SEBI already runs algorithmic-trading surveillance and TRAI regulates telecom pricing where such tools are spreading. Formal CCI protocols for information sharing would prevent fragmentation.

The agreement-centric Section 3 has served Indian competition law well against human cartels, but is ill-suited to autonomous algorithmic coordination, where harm emerges from learned interaction rather than any meeting of minds. The SPT converts those dimensions into a cumulative standard that Section 3A operationalises.

The framework is calibrated for proportionality: it targets only the narrow intersection of collusion-capable design, conducive conditions and demonstrable effects; preserves space for legitimate pricing; extends safe harbour to firms with verified safeguards; and reaches the developers best placed to prevent collusion-prone design. With the Draft Digital Competition Bill, 2024 regulating dominant platforms but leaving horizontal algorithmic collusion unaddressed, Section 3A fills a gap that will widen as adaptive pricing proliferates.

Avichal Kumar is a LL.M. student at National Law University, Delhi.