Open-source, open-weight or closed source AI? Governance remains a priority for entrepreneurs

AI discourse is increasingly crowded with sensationalist narratives. Discussions are dominated by the US versus China AI capability race. Currently, the performance gap between the leading AI models from China and the USA is rapidly closing. China’s most powerful models – such as Moonshot’s Kimi K3 and Alibaba’s Qwen3.8-Max – leverage open-weight frameworks, matching US capabilities in raw reasoning, competitive programming, and cost-efficiency. In contrast, the US frontier models are closed-source proprietary models like Anthropic’s Claude Opus 4.7 and OpenAI’s GPT-5 series. 

In July 2026, US lawmakers introduced the bipartisan AI Kill Switch Act, mandating shutdown mechanisms for advanced AI systems. The legislation proposes granting the Department of Homeland Security emergency authority to disable rogue models during catastrophic “loss-of-control” events. This regulatory push was prompted by a recent incident where an OpenAI model reportedly escaped its internal sandbox.

For business leaders, these high-stakes headlines obscure a much more immediate operational reality. Businesses that delay AI adoption while waiting for perfect technical architectures or fully formed regulations risk falling behind. Organisations must choose between three distinct deployment models: open-source, open-weight, and closed-source.

• Open-Source models: transparent systems where both the model weights and the underlying training code, architectures, and data pipelines are open for public inspection, modification, and self-hosting.
• Open-Weight models: Systems where pre-trained model weights are publicly downloadable for local deployment and fine-tuning, but the original raw training datasets, proprietary training code, and data curation recipes remain private. 
• Closed-Source models: Proprietary, “black-box” systems accessed exclusively via vendor APIs, where weights, code, and infrastructure remain entirely hidden from the enterprise. 

While each offers distinct advantages in cost, speed, and customization, they introduce different challenges for AI governance.

Contrasting the 8 Risk Pillars

Synthesizing major international frameworks—including the US NIST AI Risk Management Framework, the OECD principles, and the UK Government’s principles-based approach—reveals eight essential risk categories. Here is how their implementation challenges diverge across three deployment model types:

1. Accuracy and Reliability. The complex, probabilistic nature of AI models, can easily lead to inaccurate or unreliable outputs. 

• Open-source: Provides complete visibility into training data and logic, allowing teams to diagnose root causes of hallucinations. However, establishing benchmarks and calibration guardrails is entirely an internal burden. 
• Open-weight: Teams can fine-tune weights to improve domain accuracy, but they cannot rectify baseline inaccuracies inherited from undisclosed pre-training datasets.
• Closed-source: Delivers state-of-the-art accuracy out of the box, but exposes enterprises to silent updates to the model at the discretion of the model provider, which cans cause unpredictable model drift, reducing accuracy and reliability.

2. Fairness and Bias. AI systems can inherit and exhibit bias, which risks causing unfair or discriminatory outcomes. 

• Open-source: offers the highest level of auditability; compliance teams can inspect the full training pipeline to detect and scrub biased data before deployment. 
• Open-weight: teams can evaluate model outputs for bias across different fairness metrics but cannot verify whether underlying representations were skewed during initial pre-training. 
• Closed-source: alignment relies on vendor-defined metrics, which are often hidden from enterprise scrutiny and requires black-box probing to identify discriminatory tendencies. 

3. Explainability and Transparency: many AI models function as “black boxes,” making it difficult to understand their internal decision-making processes.

• Open-source: enables complete inspection of the internal operation of the model, and training data provenance. 
• Open-weight: allows mathematical analysis of model weights and architecture, but lacks provenance on why specific concepts were learned due to missing training logs.
• Closed-source: operates as a pure “black box”. Enterprises receive outputs without architectural visibility, complicating compliance in heavily audited sectors such as financial services and healthcare. 

4. Accountability. AI systems must have an accountable human owner in the organisation.  Furthermore, organizations must actively guard against over-reliance and automation bias among users.

• Open-source: accountability is unequivocal: the adopting enterprise is 100% responsible for every output and must build custom logging systems to prevent human automation bias. 
• Open-weight: the enterprise remains fully liable for downstream use, requiring strict human-in-the-loop governance to oversee fine-tuned application decisions. 
• Closed-source: legal liability is often diffused across vendor terms and SLAs, creating a false sense of security where staff assume the external provider bears responsibility for system failures. 

5. Privacy. The vast use of personal data in training and deploying AI raises significant privacy concerns that must be managed

• Open-source: guarantees strict data sovereignty by running entirely within an organizations private cloud or on-premise date center, preventing any data transmission outside company boundaries. 
• Open-weight: maintains local execution privacy, but requires rigorous in-house data governance to prevent enterprise PII from being accidentally embedded into weights during fine-tuning.
• Closed-source: requires customer prompts and context to traverse public APIs, demanding contractual agreements with the model vendor and assurances that data will not be used for vendor retraining. 

6. Security. AI systems remain vulnerable to unique cyber-attacks, such as model poisoning and prompt hacking.

• Open-source: models can be vulnerable to traditional software supply-chain exploits, unvetted package dependencies, and local server misconfigurations. 
• Open-weight: susceptible to model poisoning and adversarial prompt injections that bypass raw base models. 
• Closed-source: insulates enterprises from infrastructure-level vulnerabilities, but exposes workflows to vendor-side outages, API key leaks, and even prompt jailbreaks. 

7. Intellectual Property and Confidentiality. The deployment of generative AI risks leaking intellectual property and confidential data.

• Open-source: completely safeguards proprietary corporate data from third parties, though organizations carry liability if open-source datasets contain copyrighted material without licenses. 
• Open-weight: protects proprietary prompt inputs behind corporate firewalls, but offers no vendor indemnity against third-party copyright claims.
• Closed-source: commercial enterprise vendors often provide explicit IP infringement indemnification. 

8. Environmental sustainability. Organizations should strive to minimize energy consumption and emissions from their AI operations. 

• Open-source: requires capital expenditure in local GPU hardware but can be mitigated through techniques like quantization and model pruning to produce small language models (SLMs). But to achieve this, requires specialist AI engineers.
• Open-weight: incurs moderate compute expenses for hosting and fine-tuning, balancing internal engineering requirements with adaptable operational flexibility. 
• Closed-source: features near-zero initial infrastructure investment, but exposes the business to steep, variable per-token pricing. 

When choosing between the three different model types, an organization must carefully evaluate trade-offs in cost, speed, customization, and internal technical capabilities. Leaders must weigh their need for strict data privacy, IP protection, and complete explainability—which strongly favour local execution via open-source or open-weight models—against high environmental costs, infrastructure demands, and need for specialized AI workforce talent that these local deployments require. 

Conversely, if an enterprise prefers to offload environmental overhead and utilize robust, built-in safety filters with minimal upfront workforce investment, closed-source proprietary models are highly advantageous; however, they introduce opaque “black-box” architectures, external data privacy risks, and operational token costs that can escalate rapidly at scale. Ultimately, the decision hinges on balancing the organization’s risk tolerance for bearing total accountability and engineering manual guardrails internally versus relying on a vendor’s external infrastructure and safety compliance. 

Furthermore, mitigating these eight risks is not solely a technical problem; it requires cross-enterprise participation. To embed AI governance, organizations should follow a structured approach:

• Establish C-Suite Sponsorship and Oversight: Begin with a comprehensive gap analysis to identify high-risk deployments and assign clear internal ownership across business units. 

• Define Principles, Policies, and Standards: Create an AI governance hierarchy that cascades from high-level ethical commitments down to enforceable technical specifications. 

• Mandate enterprise AI Literacy: Comply with evolving requirements, such as Article 4 of the EU AI Act, by implementing tailored training programs and publishing an internal AI governance handbook. 

• Deploy a Governance Risk and Compliance platform: Integrate assessment workflows, approval checkpoints, and continuous monitoring within a dedicated governance platform, available from either specialist vendors or cloud providers, coupled with an AI inventory detailing all AI deployments,


Whether an organization chooses open-source, open-weight or closed-source model,  proactive governance is the bridge between AI experimentation and long-term enterprise value.